Centre of Natural Language Processing
Faculty of Informatics
Masaryk University
Brno
Computer Processing of
Czech Syntax and Semantics
Aleš Horák
Brno, 2008
Aleš Horák
Faculty of Informatics, Masaryk University
Centre of Natural Language Processing (NLP Centre)
Botanická 68a
CZ-602 00 Brno, Czech Republic
E-mail: hales@fi.muni.cz
Reviewed by Karel Pala, Masaryk University, Czech Republic
This work is subject to copyright. All rights are reserved, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, re-use of
illustrations, recitation, broadcasting, reproduction on microfilms or in any other way, and
storage in data banks. Duplication of this publication or parts thereof is permitted only
under the provisions of the Czech Copyright Law, in its current version, and permission for
use must always be obtained from NLP Centre, Faculty of Informatics, Masaryk University.
Violations are liable for prosecution under the Czech Copyright Law.
Copyright c NLP Centre, Faculty of Informatics, Masaryk University, 2008
ISBN: 978-80-7399-375-7
Preface
This book presents the results of research obtained during the course of
a number of natural language processing projects that were led by Aleš
Horák in the Centre of Natural Language Processing (aka NLP Centre
or NLP laboratory), Faculty of Informatics, Masaryk University in Brno.
As such, the presented results are based on the team work of researchers
as well as students who were directly participating in the projects.
The whole text offers a survey of sophisticated research methods concentrating
on two complex levels of natural language (NL) processing,
namely those of syntax and semantics. However, we do not try to cover
all approaches in these areas – we focus on rule-based introspective methods
with an encapsulation of empirical paradigms in the form of Figures
of Merits (FOMs) of particular syntactic and semantic phenomena.
As a basis for many NLP research projects, we have developed a number
of advanced natural language processing tools and language resources.
In the second chapter, we offer a detailed description covering three years
of the development of VerbaLex, a large lexicon of Czech verb valencies
in the form of complex valency frames. This part is then followed by
the presentation of specific software developed for working with this as
well as other language resources. These tools are VisDic, DEBVisDic,
DEBDict, PRALED and others. These tools are used by project teams
all over the world.
The next chapter presents the latest development of the syntactic
analyser synt that has been under development in the NLP Centre for
several years. Besides the comprehensive description of synt inside and
formats used, we also provide a comparison with several other natural
language parsers, in which we show that the synt qualities are at least
comparable to the best current parsers.
iii
PREFACE PREFACE
The fourth chapter outlines the advances made in the Normal Translation
Algorithm (NTA) from [Hor02]. It describes the methods and
techniques aimed at an automatic translation from a NL sentence to its
meaning expressed as a construction in Transparent Intensional Logic.
The description is not complete, yet, but we have concentrated on selected
phenomena where we offer sample solutions or even prototypical
implementations.
The last chapter gives details of a project that concentrates on intelligent
methods for increasing the reliability of electrical networks. One
its part involves the development of a human-machine communication
framework for dialogues about the specific knowledge domain of electrical
power systems (EPS). The task of another project part is the development
of a multi-agent system for representing the EPS processes.
These allow simulating different configurations of an EPS setup with an
automatic computation of the economic aspects of the system failures.
Acknowledgements
For more than 10 years, all research in the NLP Centre has been driven
and inspired by its head, Karel Pala. For this reason above all others
he is the first person I would like to thank for his invaluable help and
support. In addition, my thanks for cooperation on this work go to the
following NLP Centre members (in alphabetical order): Pavel Cenek,
Andrej Gardoň, Jiří Golembiovský, Dana Hlaváčková, Vladimír Kadlec,
Vojtěch Kovář, Martin Kudlej, Petr Pelán, Martin Povolný, Miroslav
Prýmek, Adam Rambousek, Pavel Rychlý, Lukáš Svoboda, Marek Veber
and Radek Vykydal, as well as co-authors of project publications from
other research teams: Marie Duží, Tomáš Holan, Pavel Materna, Pavel
Smrž, and Piek Vossen.
The financial support of this work was provided by the Faculty of Informatics,
Masaryk University, Brno and by the following grant projects:
• the EU project Balkanet IST-2000-29388,
• the Czech Science Foundation projects 201/05/2781 and 405/03/
0913,
• the projects 1ET100300414, 1ET100300419 and 1ET200610406 of
the Grant Agency of the Academy of Sciences of CR,
• the Ministry of Education of the Czech Republic within the Re-
iv
PREFACE PREFACE
search Intent CEZ:J07/98:143300003,
• the Ministry of Education of the Czech Republic within the Center
of basic research LC536, and
• the Ministry of Education of the Czech Republic in the National
Research Programme II project 2C06009.
v
PREFACE PREFACE
vi
Contents
Preface iii
1 Introduction 1
2 New Language Resources and Tools 7
2.1 VerbaLex – New Comprehensive Lexicon of Verb Valencies
for Czech . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.1.1 Linguistic Requirements for the VerbaLex Format 11
2.1.2 Semantic Roles . . . . . . . . . . . . . . . . . . . . 14
2.1.3 The Implementation of Editing and Exporting Tools 18
2.1.4 Application of VerbaLex in Syntactic Analysis . . 23
2.2 VisDic – Off-line WordNet Editor . . . . . . . . . . . . . . 25
2.2.1 Basic Functionality . . . . . . . . . . . . . . . . . . 25
2.2.2 Advanced Functionality . . . . . . . . . . . . . . . 28
2.2.3 XML Configuration . . . . . . . . . . . . . . . . . 31
2.3 DEBVisDic and other DEB Platform Applications . . . . 39
2.3.1 The Features of the Platform for Lexicographers’
Tools . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.3.2 Assets of the DEB Platform . . . . . . . . . . . . . 43
2.3.3 The DEB Administration Interface . . . . . . . . . 44
2.3.4 How To Make a Sample Dictionary . . . . . . . . . 50
2.3.5 Usage Variability – The Users’ Interfaces . . . . . 54
2.4 Future Work on the Language Resources and NLP Tools . 66
3 synt – Czech Syntax Analyzer 69
3.1 The Grammar Development Process . . . . . . . . . . . . 73
vii
CONTENTS CONTENTS
3.1.1 Grammar Development Workbench . . . . . . . . . 73
3.2 New Meta-grammar Constructs in synt . . . . . . . . . . 76
3.2.1 The Meta-grammar Design . . . . . . . . . . . . . 77
3.2.2 The Parsing Algorithm . . . . . . . . . . . . . . . 82
3.2.3 Evaluation of Contextual Constraints . . . . . . . 83
3.2.4 The synt Parser Implementation . . . . . . . . . . 84
3.3 Best Analysis Selection – a Supervised Construction of
Pruning Constraints . . . . . . . . . . . . . . . . . . . . . 87
3.4 Parsing with Verb Frame Information . . . . . . . . . . . 89
3.4.1 Automatic Extraction of Verb Frames from the Packed
Shared Forest . . . . . . . . . . . . . . . . . . . 91
3.4.2 Examples . . . . . . . . . . . . . . . . . . . . . . . 92
3.5 The Beautified Chart Method – Pruning Technique Based
on Linguistic Adequacy . . . . . . . . . . . . . . . . . . . 94
3.5.1 Beautified Trees . . . . . . . . . . . . . . . . . . . 95
3.5.2 The Previous Estimate of the Effect of the Beautified
Chart Method . . . . . . . . . . . . . . . . . . 97
3.5.3 The Beautified Chart Algorithm . . . . . . . . . . 101
3.5.4 The Beautified Chart Results . . . . . . . . . . . . 105
3.6 Parser Comparison Experiments . . . . . . . . . . . . . . 106
3.6.1 synt and Moore’s parser . . . . . . . . . . . . . . . 106
3.6.2 Phrasal synt Compared with Dependency Parsers 107
3.7 Further Development of synt . . . . . . . . . . . . . . . . 116
4 Transparent Intensional Logic as a Way to Semantics 119
4.1 Overview of the Transparent Intensional Logic . . . . . . 121
4.1.1 TIL Types . . . . . . . . . . . . . . . . . . . . . . 122
4.2 The Logical Analysis of a Sentence . . . . . . . . . . . . . 124
4.2.1 Verb Frame Analysis . . . . . . . . . . . . . . . . . 126
4.2.2 The Sentence Analysis . . . . . . . . . . . . . . . . 130
4.3 Sentence Logical Analysis Using Complex Valency Frames 134
4.3.1 Examples of Logical Analysis . . . . . . . . . . . . 136
4.4 TIL Knowledge Base Representation . . . . . . . . . . . . 142
4.4.1 Knowledge Base Implementation – the Dolphin System
. . . . . . . . . . . . . . . . . . . . . . . . . . 144
4.5 Experiment of Using TIL in a Simulation System Easel . 153
4.5.1 Using TIL and Easel in Applications . . . . . . . . 154
4.6 Long Way to Full Natural Language Semantics . . . . . . 156
viii
CONTENTS CONTENTS
5 Application in Dialogues – the Electrical Power Systems
Simulation 159
5.1 Multi-Agent Framework for EPS Simulation and Monitoring163
5.1.1 The Rice System Architecture . . . . . . . . . . . . 165
5.2 Human-machine Dialogues with the Rice System . . . . . 179
5.2.1 Building a Specialized Corpus . . . . . . . . . . . . 180
5.2.2 Morphological Tagging . . . . . . . . . . . . . . . . 182
5.2.3 Syntactic Analysis of the Domain Texts . . . . . . 182
5.2.4 Designing an Intelligent Dialogue Interface . . . . 184
5.3 Rice Usage Scenarios . . . . . . . . . . . . . . . . . . . . . 188
5.3.1 Example Nodes Implementation . . . . . . . . . . 192
5.4 Modelling of Economic Aspects of a Power System Failure 196
5.4.1 Network Parts . . . . . . . . . . . . . . . . . . . . 199
5.4.2 The Agents Implementation . . . . . . . . . . . . . 201
5.4.3 The Simulation Run . . . . . . . . . . . . . . . . . 204
5.5 Future Work on the Rice System . . . . . . . . . . . . . . 206
6 Conclusions and Future Directions 207
Bibliography 211
Annotation 227
ix
CONTENTS CONTENTS
x
Chapter 1
Introduction
The Natural Language Processing Laboratory was established at the
Faculty of Informatics, Masaryk University in 1997 within the project
VS97028 of the Ministry of Education of the Czech Republic. The founding
members of the laboratory at that time were Karel Pala, Aleš Horák,
Pavel Rychlý and Pavel Smrž. In 2005 the NLP laboratory was renamed
to the Centre of Natural Language Processing. Since its beginnings the
NLP laboratory or the NLP Centre has remained true to the essential notion
of its raison d’ˆetre: – the NLP Centre is a place where about twenty
researchers and dozens of undergraduate and postgraduate students work
on research tasks from the exciting domain of the computational processing
of written and spoken natural language. In this text, we present the
details of four of the projects in which the NLP Centre has been engaged
over last five years. What all these projects have in common is the fact
that the author of this text is the leading person of the project work.
As a first project we introduce new language resources and new language
processing tools developed in the NLP Centre. The most important
of the resources described is a new lexicon of complex valency frames of
Czech, named VerbaLex. The lexicon includes all the usual verb valency
features plus additional relevant information such as verb aspect, verb
synonymity, types of use and semantic verb classes based on the VerbNet
project. An important property of VerbaLex as far as the computer
processing is concerned, is its close relationship with the widely exploited
English and Czech WordNet semantic networks.
1
1. INTRODUCTION
We also present new tools based on client/server XML database system
called DEB ii. Thanks to the versatility of the XML format used, this
system enables us to cover various applications, namely the management
of the electronic readable dictionaries, WordNet-like lexical databases as
well as ontologies for Semantic Web applications. Considerable attention
is paid to the inner workings of the DEB ii framework as well as to particular
DEB client tools especially to WordNet development tools VisDic
and DEBVisDic, which represents a well-designed and developed system
for lexical database editing and is currently employed in many national
WordNet building projects. We discuss the basic features of the tools as
well as more elaborate functions that facilitate linguistic work in multilingual
environments. We argue for the benefits the new DEB ii platform
brings to WordNet editing and to XML databases in general.
In the following text, the main features and assets of the DEB ii dictionary
writing platform are outlined, and the implementation strategies
of both server and client parts of the platform are characterized. We also
pay attention to the process of merging lexical data, particularly Czech
WordNet with VerbaLex, the list of Czech valency frames, developed separately
in XML format. We show an example of the merge which also
indicates how DEBVisDic can serve as a means for such a kind of integration.
We also point out that this type of merge can be extended to other
languages as it was done with Bulgarian and Romanian in the Balkanet
project.
The reader will also find here an overview of the current state of other
DEB ii applications that include:
• PRALED – a client for building the Czech Lexical Database,
• DEBDict – a browser for parallel viewing of several electronic dic-
tionaries,
• Cornetto – an augmented WordNet-like lexical database system,
• DEB CPA browser and editor – a client for building a database of
verb patterns as they derive from corpora (called corpus patterns),
• DEB TEDI tool – an application for building specialized terminological
dictionaries.
For each of these DEB clients we describe and demonstrate their main
features and functionality.
2
1. INTRODUCTION
The third chapter presents a survey of the latest development of the
Czech sentence parsing system synt. The presented system uses the
meta-grammar formalism, which allows us to define the grammar with
a sustainable number of meta-rules. At the same time, these meta-rules
can be automatically translated into rules for efficient and fast headdriven
chart parsing and supplemented with an evaluation of additional
contextual constraints. In general, the system represents a rule-based
approach to syntactic analysis, however, within the parsing process synt
uses empirical parameters of the input lexical items in the form of socalled
Figures of Merits (FOMs). In this respect, the system may be
viewed as a combination of the rule-based approach and stochastic approaches
involved within most of the Czech language parsing systems
developed by members of the research group of Jan Hajič [HH07].
The text includes a comprehensive description of the meta-grammar
constructs in synt as well as actual running times of the system tested
on corpus data. The lexicographer’s environment, the so-called Grammar
Development Workbench (GDW), is integrated with synt into one
system that allows a team of experts (computational linguists as well as
programmers) to cooperate on the development of a grammar covering
all frequent Czech language phenomena.
Besides the description of the synt system, we illustrate the process
of the meta-grammar development. One of the first phases is formed
by constructing corpus data for testing. We demonstrate the exploitation
of the corpus on testing a method for detecting the Best Analysis
Selection with the results of testing the synt analysis on a Czech corpus.
The section about the Best Analysis Selection discusses methods
that enhance the algorithm determining “the best” parsing tree from the
output of natural language syntactic analysis. It presents a method for
pruning redundant parse trees based on the information obtained from a
dependency tree-bank corpus.
The VerbaLex valency lexicon from the second chapter is exploited in
the syntactic parsing process. The description of the verb frame extraction
algorithm and the measured results of running its implementation on
a newspaper corpus is displayed as one of the language specific features
used in the tree ranking algorithm, which is a crucial part of the synt
mode of analysis.
The effectiveness of the enhanced parser is demonstrated by results of
two inter-system parser comparison experiments. The first tests were run
3
1. INTRODUCTION
on the standard evaluation grammars, namely ATIS, CT and PT, where
the synt system outperforms the referential implementations. The second
experiment compared the effectiveness of real text parsers of Czech
based on completely different approaches – stochastic parsers that provide
dependency trees as their outputs, and the meta-grammar parser
that generates a resulting chart structure representing a packed forest of
phrasal derivation trees. We describe and formulate the main questions
and problems accompanying such experiments, try to offer answers to
these questions and finally display factual results of the tests as measured
on 10 thousand Czech sentences.
In the fourth chapter, we describe the extended type hierarchy of
the Transparent Intensional Logic (TIL [Tic04]) as a higher order logic
theory. We also present the basic ideas of TIL constructions as a suitable
natural language knowledge representation.
TIL is a logic system, designed for representing the meaning of natural
language expressions. The system is built on a typed λ-calculus logic
with a hierarchy of types. It was created as a parallel to Montague’s
logic [Mon74], however TIL is more capable of describing natural language
semantics while retaining the simplicity of the basic idea. Moreover,
the inference rules for TIL are well defined, thus enabling us to
use constructions as an instrument for representing sentence meaning in
knowledge base systems. The connection between a construction and the
constructed object is fact-independent and is driven by the mechanism
of typed λ-calculus. Constructions carry information about relations between
the elementary parts of language expression objects.
TIL was introduced by Pavel Tichý [Tic88] with the purpose of overcoming
paradoxes arising from other modern logical systems (first order
predicate logic as well as intensional logics [Mon74]). A short summary
of the advantages of TIL over Montague’s dynamic logic can be found
in [Hor02]. TIL is well suited to handle the difficult language phenomena
such as temporal relations, (hyper)intensionality and propositional attitudes.
The techniques described in this text are part of the long-term
development of the Normal Translation Algorithm aimed at the automatic
translation of natural language sentences into TIL constructions.
We describe methods for exploiting the VerbaLex valency frames lexicon
in relation to the transparent intensional logic. We examine the relations
between complex valency frames (CVFs) and TIL constructions
of predicate-argument structures and discuss the procedure of automatic
4
1. INTRODUCTION
acquisition of the verbal object constructions.
At the end of the chapter, the design of a newly developed Dolphin
system for the effective implementation of a knowledge base and basic
question answering based on the transparent intensional logic is explained.
We will introduce the database acting as a knowledge base for
inference in TIL. The time aspect of the truth value of propositions will
be included and the basic “thinking” capabilities of the Dolphin system
will be exemplified.
We also present an experiment in which the purely logically oriented
type system of TIL is compared with the property-based types of the
Easel [Fis99] world fact simulation language. The discussion is also
oriented towards the possibility of applications combining the two ap-
proaches.
The last of the projects we present deals with the role of biologically
motivated emergent systems and intelligent agents in the simulation of
electric power networks. The main aim is to provide a platform for analyzing
the databases of failures of power systems in the Czech Republic
(and a part of Slovakia) and to point out where the potential weak points
are. The developed system, called Rice, is designed for simulating electrical
power system processes. The system is based on the multi-agent
approach, which allows a unique versatility of the design and development
of the particular power system network. The applications of the
system aim at off-line analysis and prediction of power system failure.
The communication among the defined agents is based on standards in
multi-agent systems – the communication protocols CORBA (Common
Object Request Broker Architecture) and KQML (Knowledge Query and
Manipulation Language). The system itself economizes the open source
implementation of the protocols.
The whole system is able to perform an active simulation of the energy
flow in a power system and its visualization. We take into account a
possible future replacement of any particular agent with an on-line power
equipment monitoring facility (ad-hoc sensors), which allows the whole
power system in real time to be monitored..
The latest developed features in the Rice system allow users to capture
the behaviour of dynamic emergent networks. To fulfil this requirement,
it has to cope with the complexity of local and global changes in the network
characteristics including such basic ones as the network topology is.
The text also includes several examples of typical power network com-
5
1. INTRODUCTION
ponents and the definitions of their behaviour in the emergent network
environment.
We further present an explication of the natural language processing
tasks of the NLP Centre in the design and development of a natural language
dialogue interface for querying large databases with temporal data
about electrical power network failures. The implementation of such a dialogue
interface includes the creation and preparation of several auxiliary
resources that are required for natural language processing of texts over
this specific domain. In this text, we describe the process and statistical
results of the creation of a corpus of electrical power networks texts consisting
of more than 1 million positions. We also offer preliminary results
of the syntactic analysis of the specialized corpus data and describe the
problems of morphological and syntactic analysis of such domain specific
texts.
6
Chapter 2
New Language Resources
and Tools
Each long term natural language processing (NLP) project needs to work
with a firm basis of representative language resources and appropriate
tools for their processing. The NLP Centre at the Masaryk University
in Brno is taking part in research in all areas of the NLP field with
the “handicap” of concentrating on the Czech language, in contrast to
the most frequent languages like English, German or French. The situation
with the Czech language resources is, however, getting better
during last years with the valuable corpus resources at the beginning
(DESAM [PRS97], the Czech National Corpus [CNC06, KKK00] or the
Prague Dependency Treebank [Haj04a]) and the consecutive specialized
resources including the verb frame lexicons VALLEX [ŽL04] and VerbaLex
(see the Section 2.1).
In the first section of this chapter, we summarize the three years
of building the new VerbaLex lexicon of Czech verb frames containing
more than 10 000 verbs. The features of the lexicon are designed to bring
important semantic information to computer processing of predicate constructions
in running texts in the form of complex valency frames (CVFs).
The most notable attributes of CVFs include synset (synonymical set) organization,
two-level semantic labels with linkage to the Princeton Word-
7
2. NEW LANGUAGE RESOURCES AND TOOLS
Net and EuroWordNet1
hierarchy and surface verb frame patterns used
for automatic syntactic analysis.
In the area of lexical databases, ontologies and common sense knowledge
resources, the Princeton WordNet [Mil90] became one of the most
popular ones. It is currently used in many areas of natural language
processing such as information retrieval, automatic summarization, document
categorization, question answering, machine translation etc. To
integrate into the applications, many researchers work with the Princeton
database and transform data to their own proprietary formats.
The Princeton team also developed a data browser for WordNet which
could be downloaded together with English data from the WordNet project
web page [WN07] both for Windows and UNIX platform. Currently,
this browser is replaced with a purely web access application. No WordNet
editing tools are provided as the only instruments for majority of the
lexicographic work in Princeton are standard text editors. The consistency
of data is not therefore checked during the editing process itself, it
is postponed to later phases.
Year by year the number of Princeton WordNet derivatives and WordNet-inspired
initiatives increased. In 1998–1999 the EU project EuroWordNet
1 and 2 [Eur99] took place, in which multilingual approach
dominated and WordNets for 8 European languages, particularly for English,
Dutch, Italian, Spanish, French, German, Czech and Estonian, were
developed. The Interlingual Index (ILI), the Top Ontology, set of Base
Concepts and a set of Internal Language Relations were introduced as
well [Vos98]. These changes also led to the design and development of a
new database engine for EuroWordNet and it resulted in the editing and
browsing tool called Polaris [Lou98].
In 2001–2004 the EU project Balkanet [Bal04] was launched which
can be viewed as a continuation of the EuroWordNet project. It was
conceived as multilingual, as well, and within its framework WordNets
for 6 languages were developed or augmented, particularly for Greek,
Turkish, Romanian, Bulgarian, Serbian and Czech. Before Balkanet had
started it was already obvious that the Polaris tool had no future because
its development had been closed and as a licensed software product (by
Lernout and Hauspie) it had been rather expensive for most of the research
institutions involved (typically universities). Moreover, the system
1see further text for details
8
2. NEW LANGUAGE RESOURCES AND TOOLS
had been provided only for MS Windows platform. That is why a specialized
open source software system VisDic has been developed by MU
NLP Centre during the work within the Balkanet project. VisDic was
designed and implemented as a highly configurable multiplatform tool
with easy to use interface for working not only with WordNets, but with
general dictionaries using a variant of XML schema for the dictionary
entries. We present the details of this tool in the Section 2.2.
In June 2000, the Global WordNet Association (GWA, see the web
site [GWA07]) was established by Piek Vossen and Christiane Fellbaum.
The purpose of this association is to “provide a platform for discussing,
sharing and connecting WordNets for all languages in the world.” One of
the most important actions of GWA is the Global WordNet Conference
(GWC) that is being held every two years on different places all over
the world. The second GWC was organized by the MU NLP Centre in
Brno and the NLP Centre members are actively participating in GWA
plans and activities. A new idea that was born during the third GWC in
Korea is called the Global WordNet Grid with the purpose of providing
a free network of smaller (at the beginning) WordNets linked together
through Interlingual Ontology (as opposed to Interlingual Index from
EuroWordNet). The Grid preparation is currently just starting and the
MU NLP Centre is developing its software background.
The growing need to handle various lexical resources that take the
form of dictionaries, semantic networks, ontologies, valency lexicons, or
FrameNets is the cause why researchers seek for software systems that are
able to store dictionary-like data using XML as the core element. Many
dictionary publishing houses operate large systems with the complex
functionality of so called lexicographic stations that manipulate XML
(DPS Longman [McN03]) and several companies offer dictionary writing
programs of different complexity (TshwaneLex [JdS04], iLEX [Erl04] or
Field Linguist’s Toolbox [Too07]). However, these and similar tools are
not always able to efficiently manipulate resources obtained from datadriven
NLP applications. Therefore, they cannot provide a universal
environment for lexical database management as well as semantic networks
and ontologies. They often represent rather large systems that are
quite complex which is not always an advantage. And, last but not least,
some of them are rather expensive. That is why we decided to build a
development framework on which the individual clients can be built –
this solution is modular and flexible since the clients can be adapted for
9
2.1 2. NEW LANGUAGE RESOURCES AND TOOLS
the particular purpose in a short time. The description of this development
platform, called DEB ii, forms the content of the third part of this
chapter.
The contents of this chapter is an extension of the previous work
published in [HS03, HS04, HH05a, HH05b, HPRP06, HP06a, PH06a,
HPRR06, HR07, HP07, HVR08].
2.1 VerbaLex – New Comprehensive Lexicon
of Verb Valencies for Czech
The beginnings of building the verb valency frame dictionary at the MU
NLP Centre dates back to 1997, when Karel Pala prepared the first version
of a verb valency dictionary with 15 000 entries [PŠ97]. Since then,
the dictionary, denoted as Brief, has undergone a long development and
has been used in various tools from semantic classification to syntactic
analysis of Czech sentence [SH98]. The data in this dictionary can be
entered in several mutually convertible formats:
brief format:
jíst2
<v>hTc4,hTc4-hTc6r{na}, hTc4-hTc7
verbose format: display format:
jíst2
= co jíst něco
= co & na čem jíst něco na něčem
= co & čím jíst něco něčím
The Brief dictionary contains about 15 000 verbs with 50 000 verb valency
frames, thus making it an invaluable language resource with high
coverage. However, the different verb senses are not distinguished here.
Another advance in the Czech verb valency processing came during
the work on the Czech WordNet within the Balkanet project (see [Bal04]).
The Czech WordNet has been supplemented with a new language resource,
Czech WordNet valency frames dictionary. The new acquisition
2‘jíst’=‘to eat’, ‘co’=‘what’, ‘na čem’=‘on what’, ‘čím’=‘(with) what’,
‘něco’=‘something’, ‘na něčem’=‘on something’, ‘něčím’=‘(with) something’
10
2. NEW LANGUAGE RESOURCES AND TOOLS 2.1.1
of this dictionary were the semantic roles and links to the Czech WordNet
semantic network.
During the work on enhancing the list and adding new entries into
it, we have come to the need of comparing the quality and features of
the list with the valency lexicon of Czech verbs denoted as VALLEX
1.0 [SLŽ02] that was created independently of the Czech WordNet verb
frames. Based on these three verb frame resources, we have designed new
format of a verb frame entry named the complex valency frame (CVF).
The resulting lexicon of Czech verbs CVFs, named VerbaLex, contains
information useful for automatic computer processing of verb frames with
the linguistic background.
The VerbaLex dictionary is being actively developed, checked and
supplemented with new data since 2005. The coordination of the work
of 4 IT developers and 15 linguists is directed by Aleš Horák with Dana
Hlaváčková as the head of the linguistic team. Currently, VerbaLex contains
10 782 verb lemmata which, when gathered in synonymic groups,
share 28 566 verb frames.
2.1.1 Linguistic Requirements for the VerbaLex For-
mat
In this section, we present the substantiation of the main differences
between VerbaLex and VALLEX 1.0 valency frames notation.
VerbaLex differs from VALLEX 1.0 in augmentation of the original
format, detailed differentiation of valency frames and above all two-level
semantic roles.3
The headwords in VerbaLex are formed with lemmata in
a synonymic relation followed by their sense numbers (standard Princeton
WordNet notation, such an expression is denoted as a WordNet literal).
The lexical units in WordNet are organized into synsets (sets of
synonyms) arranged in the hierarchy of word meanings (hyper-hyponymic
relations). The standard definition of synonymy says that two synonymic
words can be always substituted in the context. However, the synonymy
in synsets is understood like very close sense affinity of given words, the
substitution rule cannot be applied in all cases here. In VALLEX 1.0, a
headword is one lemma, possibly two or more lemmata in case of lemma
3for more details about the VerbaLex semantic roles see the Section 2.1.2
11
2.1.1 2. NEW LANGUAGE RESOURCES AND TOOLS
Lemma variants:
Princeton WordNet – plan:2
Definition: make plans for something
VALLEX 1.0: vymyslet2 / vymyslit2
VerbaLex: vymyslet:1, vymyslit:1, naplánovat:3
Word entries:
Princeton WordNet – arrive:1, get:5, come:2
Definition: reach a destination; arrive by movement or
progress
VALLEX 1.0: dojít1
VerbaLex: dojít:1, dorazit:1, dostat se:1, přicestovat:1,
přijet:1, přijít:1
Figure 2.1: Examples of verb frame entry heads for verbs with lemma
variants and for synonymic verbs.
variants.4
Lemma variants in VerbaLex are considered as independent
lemmata and they are distinguished by their WordNet sense numbers.
An example of two verb frame entries in VALLEX 1.0 and VerbaLex is
displayed in the Figure 2.1.
In VerbaLex, each word entry includes an information about the verb
aspect (perfective – pf., imperfective – impf. or both aspects – biasp.).
VerbaLex valency frames are enriched with aspect differentiations for
examples containing the verb used with the given valency frame. This is
important in case of synonymic lemmata with different aspect:
Princeton WordNet – wade:1
Definition: walk (through relatively shallow water)
VerbaLex: brodit se:2impf., přebrodit se:1pf.
frame: AG<person:1>obl
who nom VERB
SUBS<substance:1>obl
(through+)what ins
example: přebrodil se blátempf. / he wade through mud
4the lemmata with small phoneme alternation (in Czech) that are interchangeable
in any context without any change of the meaning – bydlet/bydlit = to live
(somewhere).
12
2. NEW LANGUAGE RESOURCES AND TOOLS 2.1.1
example: brodil se pískemimpf. / he wade through sand
The constituent elements of frame entries are enriched with pronominal
terms (e.g. who, what) and the morphological case number or short
word.5
This notation allows to differentiate an animate or inanimate
agent position:6
Princeton WordNet – bump:1, knock:3
Definition: knock against with force or violence
VerbaLex: narazit:1pf. / narážet:1impf.
frame: AG<person:1>obl
who nom VERB
OBJ<object:1>obl
to+what gen,at+what acc
PART<body part:1>obl
(with+)what ins
example: I bumped to the wall with my head
frame: OBJ<vehicle:1>obl
what nom VERB
OBJ<object:1>obl
to+what gen,at+what acc
example: the car bumped to the tree
2.1.1.1 Verb Usage and Verb Classes
VerbaLex captures additional information about types of verb use and
semantic verb classes. Three types of verb use are displayed in the lexicon.
The primary usage of a verb is marked with abbreviation prim,
metaphorical or figurative use with fig and idiomatic and phraseological
use with idiom (this notation corresponds to the original VALLEX 1.0).
The assigned semantic verb classes have been adopted from the Martha
Palmer’s [DKPR98] VerbNet project. The verb classes list is based on
Beth Levin’s [Lev93] classes with more fine-grained sets of verbs.
There are 395 classes in the current development version of VerbNet,
which was provided by Martha Palmer’s team. But this number seems to
be too high for Czech verbs, therefore the list of verb classes is adapted
to the conditions of the Czech language:
Princeton WordNet – cry:2, weep:1
5for the 7 grammatical cases in Czech – nom[inative], gen[itive], dat[ive],
acc[usative], voc[ative], loc[ative] and ins[trumental]
6The prepositions in VerbaLex are stated in Czech. Here, in the presented examples,
we translate them to English.
13
2.1.2 2. NEW LANGUAGE RESOURCES AND TOOLS
Definition: shed tears because of sadness, rage, or pain
VerbaLex: brečet:1, plakat:1, ronit:1
class: nonverbal expression-40.2
Princeton WordNet – take care:2, mind:3
Definition: be in charge of or deal with
VerbaLex: dbát:2, starat se:2, pečovat:3
class: care-86
Princeton WordNet – be:11, live:5
Definition: have life, be alive
VerbaLex: žít:1, být:2, existovat:3
class: exist-47
2.1.2 Semantic Roles
Semantic role annotation is usually based on the appropriate inventories
of labels for semantic roles (deep cases, arguments of verbs, functors,
actants) describing argument predicate structure of verbs. It can be
observed that different inventories are exploited in different projects (e.g.
VALLEX [ŽL04], VerbNet [KDP00], FrameNet [FBS04], Salsa [BPGR06],
CPA [Han04] or VerbaLex here).
The idea of semantic roles in VerbaLex has come from the specification
of needs of the Czech syntactic analysis – we need a technique for
distinguishing sentence constituents as a) obligatory,7
b) typical (for the
purpose of syntactic tree ranking,8
or c) forbidden, for the sake of tree
pruning.
These considerations led us to the design of the inventory of twolevel
labels which are presently exploited for annotating semantic roles in
VerbaLex. VerbaLex has thus introduced a different concept of semantic
roles than in the VALLEX 1.0 project.9
The functors used in VALLEX
1.0 valency frames seem to be too general and they do not allow to
distinguish different senses of verbs according to the WordNet style. We
suppose that a more specific subcategorization of the semantic role tags
is necessary for the needs we have defined above.
7with regard to the valency of another sentence constituent
8see the Sections 3.2.2 and 3.2.3
9semantic roles are denoted as functors in VALLEX 1.0
14
2. NEW LANGUAGE RESOURCES AND TOOLS 2.1.2
The first level of VerbaLex semantic labels contains the main semantic
roles proposed on the basis of 1st
and 2nd
Order Entities from the EuroWordNet
Top Ontology [VB+
98]. On the second level, we use selected
particular literals (lexical units) from the set of Princeton WordNet Base
Concepts with relevant sense numbers. We can thus specify groups of
words (hyponyms of these literals) replenishable to valency frames. This
concept allows us to specify valency frames notation with large degree
of sense differentiability. The motivation for this choice is based on the
fact that the Princeton WordNet has a hierarchical structure which covers
about 110 000 English lexical units (synsets). It is then possible to
use general labels corresponding to selected top and middle nodes and go
down the hyperonymy/hyponymy (H/H) tree until the particular synset
is found or matched. This allows us to see what is the semantic structure
of the analyzed sentences using their respective valency frames. The
nodes that we have to traverse when going down the H/H tree at the
same time form a sequence of the semantic features which characterize
meaning of the lexical unit fitting into a particular valency frame. These
sequences can be interpreted as quite detailed selectional restrictions.
Currently, we use about 40 1st
-level and 200 2nd
-level semantic roles in
VerbaLex.
For example the literal writing implement:1 is a hypernym for any
implement that is used to write.
Princeton WordNet – draw:6
Definition: represent by making a drawing of, as with a
pencil, chalk, etc. on a surface
VerbaLex: kreslit:1, malovat:1
frame: AG<person:1>obl
who nom VERB
ART<creation:2>obl
what acc
INS<writing implement:1>obl
(with+)what ins
example: my sister draws a picture with coloured pencils,
the famous artist was drawing his painting only with
charcoal
The left-side valency position is most frequently occupied by the semantic
role AG, an agent. The agent position in a valency frame is understood
as a very general semantic role (functor ACT) in VALLEX 1.0. This label
does not allow to distinguish various types of action cause. Two level
15
2.1.2 2. NEW LANGUAGE RESOURCES AND TOOLS
semantic role labels in VerbaLex are able to define cause of action quite
precisely. The main semantic role AG is completed by an adequate literal
depending on the verb sense and valency frame. Thus, we can identify
whether the agent is a person AG<person:1>, an animal AG<animal:1>,
a group of people AG<group:1>, an institution AG<institution:1> or
a machine AG<machine:1>. For some verbs with a very specific sense,
hyponyms of these literals are used. For example:
Princeton WordNet – sugar:1, saccharify:1
Definition: sweeten with sugar
VerbaLex: sladit:4, osladit:1, pocukrovat:1
frame: AG<person:1>obl
who nom VERB SUBS<food:1>obl
what acc
SUBS<sugar:1>obl
(with+)what ins
example: sugar your tea with brown sugar
In VALLEX 1.0, each valency frame starts always with functor ACT. In
our opinion, it is useful to differentiate the sense of the left-side valency
position (subject position) in more detail. According to our definition of
agent AG (somebody or something doing something actively) this position
may be also occupied by other semantic roles. The subject position can
contain objects OBJ, substances SUBS or a semantic role denoting abstract
concepts – human activity ACT, knowledge KNOW, event EVEN, information
INFO, state STATE. For example:
Princeton WordNet – follow:6, come after:1
Definition: come after in time, as a result
VerbaLex: přijít:25 / přicházet:25, následovat:4
frame: EVEN<event:1>obl
what nom VERB
EVEN<event:1>obl
(after+)what loc
example: heavy rain followed flood
Princeton WordNet – fall:3
Definition: pass suddenly and passively into a state of body
or mind
VerbaLex: zachvátit:2, zmocnit se:2
frame: STATE<state:4>obl
what nom VERB
PAT<person:1>obl
whom acc
example: he fall into a depression
16
2. NEW LANGUAGE RESOURCES AND TOOLS 2.1.2
Table 2.1: List of 1st
-level semantic roles from VerbaLex that are used in
the examples.
AG the semantic role of the animate entity that instigates or
causes the happening denoted by the verb in the clause,
we extended this definition for inanimate entity that does
sth actively (e.g. machine)
ART a man-made object taken as a whole
SUBS that which has mass and occupies space
PART a portion of a natural object, something determined in
relation to something that includes it, something less than
the whole of a human artifact
INS a device that requires skill for proper use
OBJ a tangible and visible entity; an entity that can cast a
shadow
EVEN something that happens at a given place and time
STATE the way something is with respect to its main attributes
A large number of semantic roles inspired by EuroWordNet Top Ontology
roughly correspond with the PAT functor in VALLEX 1.0. The PAT label
covers quite different senses, which can be very well identified.
In our inventory, PAT is defined as: the semantic role of an entity that
is not the agent but is directly involved in or affected by the happening
denoted by the verb in the clause (the definition of the patient:2 literal
from the Princeton WordNet).
Princeton WordNet – experience:1, undergo:2, see:21, go
through:1
Definition: go or live through
VALLEX 1.0: absolvovat2
frame: ACTobl
1 PATobl
4
VerbaLex: absolvovat:2, prožít:1 / prožívat:1
frame: AG<person:1>obl
who nom VERB
EVEN<experience:3>obl
what acc
example: he underwent difficult surgery
17
2.1.3 2. NEW LANGUAGE RESOURCES AND TOOLS
Some second level literals cannot be adopted from the Princeton WordNet
Base Concepts – especially specifications of roles considered as “classic”
deep cases. These literals (e.g. agent:6, patient:2, donor:1, addressee:1,
or beneficiary:1) do not have any hyponyms in Princeton WordNet
and cannot be substituted by any word.
For such cases, the literal person:1 is used (or another suitable literal
with large number of hyponyms, e.g. AG<person:1>, PAT<animal:1>).
This “classic” semantic roles are consistent with some functors in VALLEX
1.0 (ACT, PAT, ADDR, BEN etc.). A list of VerbaLex semantic roles
that are used in the presented examples is displayed in the Table 2.1.
2.1.2.1 Special Semantic Roles
VerbaLex describes not only the valency and semantic frames, it also
includes other relevant information about Czech verbs, such as the verb
position. In a free-word order language like Czech the position of the
verb within the verb frame is usually not strictly specified.
VerbaLex uses a special semantic role, VERB, which marks the canonic
position of the verb in its verb frame. Such default verb position is not
needed only for analysis of verb valencies, it can be also directly used
in the process of generation of Czech sentences, e.g. as an output of a
question-answering machine.
The left side of the verb position is traditionally occupied by the
sentence subject, which is also marked in most of the verb frames in
VerbaLex. However, there are some cases, where the verb frame has to
obey different rules – e.g. sentence Dalo se do deště (It started to rain)
cannot contain any subject. For the notation of such cases, VerbaLex
uses another special semantic role ISUB, an inexplicit subject.
2.1.3 The Implementation of Editing and Exporting
Tools
For the sake of editing and entry management in the newly adopted verb
valency frame format for VerbaLex, we have implemented a new set of
editing tools.
The main interactive tool for user editing of the valency dictionary,
named verbalex.sh, is based on a highly configurable multi-platform
18
2. NEW LANGUAGE RESOURCES AND TOOLS 2.1.3
Figure 2.2: The tool for editing verb valency frames dictionary in the
VerbaLex format.
editor VIM [Sch07] (see the Figure 2.2). Such approach enables a linguistic
expert to easily enter computer-parseable data in a fixed plain
text format and still, thanks to the flexible color syntax highlighting, he
or she has a full visual control of possible errors in the format.
The editing itself is not fixed to one platform, users can run the same
environment under any of the current popular computer operating systems
(VIM editor runs on nearly any platform).
The authoring tool verbalex.sh currently offers these functions to
the editing user:
• free editing of the dictionary entries
• regular expression searching in the dictionary
• template-based adding of a new verb entry or a new verb frame to
the current entry
• menu-based adding of a new semantic role to the current frame
19
2.1.3 2. NEW LANGUAGE RESOURCES AND TOOLS
• multilevel folding – hiding/displaying of valency attributes, valencies
or full valency frames
• visual marking of the current frame for further inquiry
• interactive merging of definitions from two parallel sources
Moreover, the interpreted approach of the tool makes adding new features
to the editing system easier to implement.
The plain text format edited by a human expert was inspired by the
editing format of VALLEX 1.0 [Žab05]. This text format is in further
processing transformed into an XML standard format which enables conversions
into different formats used for visual checking, searching and
presentation of the valency dictionary.
A good example of merging various lexical data is the work going
on in the NLP Centre at FI MU where the data from Czech WordNet
and Czech Valency Lexicon VerbaLex are combined together. The VerbaLex
lexicon is currently being developed separately and independently
of Czech WordNet using a particular XML format (see the Figure 2.3).
However, the entries in VerbaLex are written in form of WordNet synsets,
which enables combining the data from both these resources.
The Czech WordNet currently contains a smaller set of valency frames
in plain PCDATA format (see the Figure 2.4). The current work is directed
to merging the VerbaLex valency frames with the Czech WordNet
synset structures.
The Czech verbs for which valency frames already exist are or will
be linked to their English equivalents by means of ILI (Inter-Lingual Index).
If Czech and English verbs (synsets) are linked correctly, the deep
valency frames developed for Czech can be also valid for English (surface
valencies are obviously different since Czech is a synthetic language
whereas English is an analytic one).
The XML schema for VerbaLex also took inspiration from VALLEX
1.0, where the original schema had to be changed to suit the augmentation
of the format in VerbaLex. The changes include
• adding class attribute to frame slot tag to cover WordNet basic
concept literals
• including the WordNet word sense in the lemma tags
• shifting the verb aspect to headword lemma, which now enumerates
all the aspectual counterpart tuples. An example of such XML
substructure can be found in the Figure 2.3.
20
2. NEW LANGUAGE RESOURCES AND TOOLS 2.1.3
<word_entry>
<headword_lemmata>
<lemma ord=’1’ sense=’1’ aspect=’pf’
aspectual_counterpart_lemma=’dodávat’>dodat</lemma>
...
</headword_lemmata>
<frame_entry frame_index=’1’>
<frame_lemmata>
<lemma sense=’8’ aspect=’pf’>dát</lemma>
...
</frame_lemmata>
<synonym_lemmata>
<lemma aspect=’pf’ sense=’1’>vložit</lemma>
...
</synonym_lemmata>
<example>dok: připojili ke smlouvě své podpisy</example>
<use>prim</use>
<frame_slots>
<slot number=’1’ functor=’AG’ type=’obl’ class=’person:1’>
<form type=’direct_case’ case=’kdo1’ />
</slot>
<slot number=’2’ type=’obl’ functor=’VERB’/>
<slot number=’3’ functor=’INFO’ type=’obl’ class=’info:1’>
<form type=’direct_case’ case=’co4’ />
</slot>
<slot number=’4’ functor=’COM’ type=’obl’
class=’written communication:1’>
<form type=’prepos_case’ prepos_lemma=’k’ case=’čemu3’/>
</slot>
</frame_slots>
</frame_entry>
...
</word_entry>
Figure 2.3: An example of (a part of) an entry in the VerbaLex
XML format for the synset dát:8, vložit:1, vsunout:1, přidat:2,
připojit:1, dodat:1 (i.e. insert:1, infix:1, enter:7, introduce:6 in the
Princeton WordNet)
21
2.1.3 2. NEW LANGUAGE RESOURCES AND TOOLS
Synset: dát:8, vložit:1, vsunout:1, přidat:2, připojit:1, dodat:1
<VALENCY>
<FRAME>{dát, vložit, vsunout}
kdo1*AG(person:1)=co4*OBJ(object:1)
& do čeho2*OBJ(container:1)
</FRAME>
<FRAME>{dát, vsunout, přidat, vložit, dodat}
kdo1*AG(person:1)=co4*INFO(info:1)
& do čeho2*COM(written communication:1)
%dodal do textu nové poznámky, přidal k článku obrázek
</FRAME>
<FRAME>{dát,přidat, připojit, dodat}
kdo1*AG(person:1)=co4*INFO(info:1)
& k čemu3*COM(written communication:1)
%připojili k smlouvě své podpisy
</FRAME>
<FRAME>{přidat, připojit, dodat}
kdo1*AG(person:1)=co4*OBJ(object:1)
? k čemu3*OBJ(object:1)
%připojil hadici ke kohoutku
</FRAME>
Figure 2.4: An example of valency frames in the Czech Wordnet for
the same synset insert:1, infix:1, enter:7, introduce:6 as was
displayed in the Figure 2.3
The resulting XML structure is then transformed into various output
formats with the use of modified tools from VALLEX 1.0. The export
formats are:
• HTML with navigation among the characteristic features of the
dictionary entries,
• Postscript document for printing including page index of all verbs,
and
• PDF, which allows navigation through the document in the same
visual form as for hardcopy printing.
22
2. NEW LANGUAGE RESOURCES AND TOOLS 2.1.4
2.1.4 Application of VerbaLex in Syntactic Analysis
The design of VerbaLex verb valency lexicon was driven mainly by the
requirement to describe the verb frame (VF) features in a computer
readable form suitable for syntactic and semantic analysis. The current
CVFs10
structure contains:
• morphological and syntactic features of constituents
• two-level semantic roles
• links to Princeton WordNet and Czech WordNet hypero/hyponymic
hierarchy
• differentiation of animate/inanimate constituents
• default verb position
• verb frames linked to verb senses
• VerbNet classes of verbs.
We are currently testing the application of VerbaLex in the syntactic
analyzer synt (see the Chapter 3) that is designed for parsing real-text
sentences. For the detailed description of the verb frame extraction process
see the Section 3.4.1.
The system processing can be presented on an example sentence –
see the syntactic tree in the Figure 2.5 and the textual output of the
part of the system that implements the VFE algorithm in the Figure 2.6.
The system first identifies the verb rule constituents (nterms). Then the
corresponding groups, i.e. the actual sentence constituents that will play
the role as verb frame arguments, are extracted from the forest of values.
Groups usually do not correspond to nterms one-to-one, since they are
stored within non-terminals deeper in the forest and not directly in the
verb rule. This part of the VFE algorithm has unfortunately exponential
time complexity, however, for common sentences the depth of the verb
frame constituents is not more than three levels, so the actual running
times are usually within fractions of seconds. After the identification
of the groups, the algorithm looks for possible subjects – this is not as
easy as it may look at the first sight, since the sentence subject can
be expressed not only by a noun phrase in nominative (which is the
most frequent option in Czech), but also by e.g. prepositional phrase or
verb infinitive. If no possible subject is found, the algorithm supplies a
pronoun for an inexplicit subject with the gender corresponding to the
verb. The Clause valency list displays all possible combinations of the
10Complex Valency Frames
23
2.1.4 2. NEW LANGUAGE RESOURCES AND TOOLS
start
clause
np
ADJ
Malé
np
N
děti
part adv
ADV
špatně
V
snáší
np
ADJ
dlouhou
np
np
N
jízdu
np
N
autem
ends
’.’
.
Figure 2.5: Syntactic tree of an example input sentence “Malé děti
špatně snáší dlouhou jízdu autem.” (Small children badly withstand
long journey by car.)
verb rule schema: 3 nterms, ’#2’
nterm 1: k1gNnPc1
nterm 2: k5eAp3nPtPmIaI
nterm 3: k1gFnSc4
group 1: 0,2, npnl -> .{ left modif } np . k1gMnSc1
"malé děti"
group 2: 2,3, ADV -> .’špatně’ . k6xMeAd1
group 3: 4,7, npnl -> .{ left modif } np . k1gFnSc4
"dlouhou jízdu autem"
possible subjects: #1
Clause valency list:
snášet <v>-#2:(1)hH#1:(0)hPTc1-#3:(2)hPTc4
snášet(0) <v>#1:(1)hH-#2:(2)hPTc4
Verb valency list:
snášet <v>#2:hH-#1:hPTc4
snášet <v>#1:hPTc4
Matched valency list:
snášet(0) <v>#2:(1)hH-#1:(2)hPTc4
Figure 2.6: The output of the verb frame extraction algorithm during the
example sentence analysis.
24
2. NEW LANGUAGE RESOURCES AND TOOLS 2.2
translations of the verb arguments found into verb frame patterns. This
list is then intersected with the list of lexicon entries for the verb to obtain
the Matched valency list as a result of the VFE algorithm.
The effectiveness of the syntactic analysis with the VFE algorithm
was measured on approximately 4 000 Czech corpus sentences with the
median of 15 words per sentence and the Clause valency list contained 11
possible verb frames with the running time of 0.07 seconds per sentence.
2.2 VisDic – Off-line WordNet Editor
As the developers of Czech WordNet within EuroWordNet project we
came to the conclusion that a new tool for WordNet browsing and editing
has to be developed rather quickly. At the same time we realized that it
was necessary to look for the solution that would also support establishing
the necessary standards for WordNet like lexical (knowledge) databases.
Thus we decided to develop a new tool for WordNets based on XML data
format, which can be used for lexical databases of various sorts. The tool
is called VisDic and it has been implemented in 1999–2004 in the Natural
Language Processing Laboratory at the Faculty of Informatics, Masaryk
University for both Windows and Linux platform [PP02, HS03].
2.2.1 Basic Functionality
VisDic was developed as a tool for presentation and editing (primarily
WordNet-like) dictionary databases stored in XML format. Most of the
program behaviour and the dictionary design can be configured. With
these capabilities, we can adopt VisDic to various dictionary types—
monolingual, translational, thesaurus or generally linked WordNet lexi-
cons.
2.2.1.1 Multiple Views of Multiple WordNets
The main working window is divided into several dictionary panels. Each
panel represents a place for entering queries and browsing context of one
specified WordNet dictionary. The panels can display different WordNets
as well as multiple contexts of the same dictionary.
The contents of a panel offers, besides the query input and matching
results list, a set of overlapping notebook tabs each of which represents
25
2.2.1 2. NEW LANGUAGE RESOURCES AND TOOLS
Figure 2.7: An example of freely defined text view of a WordNet entry
one kind of view of the same entry from the list of results. The order,
the type and even the content of each notebook tab is specified by the
user in the configuration files (see the Section 2.2.3). The main types of
views are described in the following sections.
2.2.1.2 Freely Defined Text Views
The content of the Text View notebook tab is entirely built from the
user definition that follows the XML structure of the WordNet entry.
The editor can thus present an easily readable view of the entry with
highlighting important parts of the entry content (see the Figure 2.7).
2.2.1.3 Edit
The editing capabilities allow to give the user a full control over the
content and linking of each entry in the WordNet hierarchy. To prevent
a user from moving the entry as an object in the multicolored spider
web of the linkage relations, the linguist rather specifies all the links in
26
2. NEW LANGUAGE RESOURCES AND TOOLS 2.2.1
a textual dialog, where all the bindings are displayed in one place with
consistency checks after each change request.
The actual contents of the Edit notebook tab is also entirely driven
by user instructions in the configuration, where each editing field is given
a textual label and is assigned to an XML tag from the entry structure.
2.2.1.4 Tree and RevTree
WordNet dictionaries are specific with a heavy network of various kinds
of relations between the dictionary entries with the function to capture
the ontology relations of the underlying natural language.
Navigation in such environment is thus a crucial point of a successful
linguistic work with WordNet data. Since the linkage relations generally
do not need to obey any rules, that could make the resulting structure to
be an arbitrary directed graph. VisDic implements a browsing mechanism
for general graphs. The navigation process works with two interconnected
notebook tabs, which always both start at the same dictionary entry and
display its position in the graph represented as a breadth-first path trees
of all the linkage relations that lead from the entry to other entries in the
dictionary. Each of the notebook tabs displays mutually opposite linkage
relations, allowing the user to choose the direction of graph navigation
in every step.
To facilitate the orientation and to help to position the entry in the
WordNet hierarchy, the navigation also displays the path from the entry
to its top in the hyper-hyponymic relation tree (see the Figure 2.8).
For more advanced navigation the linguist may also use advanced tree
browsing techniques (described in the Section 2.2.2.3).
2.2.1.5 Query Result and External File Lists
Common actions in WordNet creation and editing often include processing
of a subset of entries based on certain criteria. VisDic offers a suitable
kind of views for this situation, which allow to prepare a notebook tab
with a list of entries matching any user specified query or a list of entries
identified by entry-IDs gathered in a plain text file.
27
2.2.2 2. NEW LANGUAGE RESOURCES AND TOOLS
Figure 2.8: The tree-like navigation in the WordNet linkage relations
graph
2.2.1.6 Plain XML View
Sometimes users need a thorough view into the data structures contained
in the dictionary entry. XML View notebook tab offers this possibility.
In this view, the user can see a graphically structured XML text, which
represents the entry structure as it is stored in the dictionary (see the
Figure 2.9).
2.2.2 Advanced Functionality
The basic functionality described in the previous section generally conforms
to any XML based dictionary. However, linguistic work specialized
to WordNet creation and editing requires some more specific and more
sophisticated functions in the editor.
28
2. NEW LANGUAGE RESOURCES AND TOOLS 2.2.2
Figure 2.9: Raw XML view of a synset entry.
2.2.2.1 Synchronization
Within the creation of a national (e.g. Czech) WordNet, which would correspond
to the English WordNet as a primary reference, one of the most
frequent operation is a lookup of a dictionary entry (synset) from one
WordNet in another dictionary. Such lookup uses either the SYNSET.ID
tag (as a direct equivalent of the entity) or one of the, so called, equivalence
tags (or attributes) defined in the configuration. An example of
such tag may be REVMAP or MAPHINT used to help the linguist to process
ambiguous link references between various versions of English WordNet.
The lookup function in VisDic can work in two modes: as an instant
(one time) lookup — the Show (by) operation, and also as a firmly established
link between two notebook tabs called the AutoLookUp (by). In
case of AutoLookUp, any move to another dictionary entry in the source
notebook tab leads to an automatic lookup of the new entry in the destination
dictionary. VisDic allows to have any acceptable combination of
AutoLookUps among all the notebook tabs.
2.2.2.2 Editing Support
The efforts of unifying national WordNets based on the English WordNet
in many cases lead to copying of synset information between different language
dictionaries. Such functionality in VisDic is split into two common
situation — either the SYNSET.ID of an existing synset is to be unified
29
2.2.2 2. NEW LANGUAGE RESOURCES AND TOOLS
with the ID of the English synset (Take key from operation) or a whole
new entry is to be copied to another dictionary (Copy entry to).
2.2.2.3 Tree Browsing
The basic navigation in related synsets (in some cases reduced to the
hypernymic and hyponymic relations tree) is supplemented with two important
WordNet operations — Topmost entries and Full expansion.
The Topmost entries operation identifies all synsets, which are (in the
tree subset of linkage relations) found as the roots of relational hierarchy,
i.e. are not hung below any other synset. This helps the linguist to
identify the level 1 entries as well as so far unfiled entries.
The Full expansion allows the user to see all possible descendants of
a selected synset in the linkage relations graph. During the operation
cycle detection techniques check any violations of tree properties in the
graph. Some relations can be also configured to be left out from the full
expansion process.11
2.2.2.4 Consistency Checks
Semi-automatic processing, which often takes part in the national WordNets
creation, as well as common human processing of the data inevitably
brings in the possibility of mistakes. The inconsistencies, which may be
revealed as duplicates, are controlled by VisDic consistency checks, which
contain
• check duplicate IDs
• check duplicate literals and senses
• check duplicate synset literals
• check duplicate synset links
These checks allow the linguist to identify the most common errors e.g. after
merging data from various sources.
2.2.2.5 Journaling
The work on a large and representative national WordNet usually employs
more than one linguist working on the data. The synchronization of
11see also Visual Definitions in the Section 2.2.3.2
30
2. NEW LANGUAGE RESOURCES AND TOOLS 2.2.3
the resulting dictionary is made possible in VisDic with the usage of
journaling.
During the work with VisDic, any change of the data is marked in
a journal file. Each journal file is specific to one dictionary and one user
at a time. Such journal file can then be “applied” to the dictionary data
and merged with the original. In this way, the simultaneous work of
several linguists can be easily interchanged with a common data source.
2.2.3 XML Configuration
Most of the functionality in the VisDic WordNet editor can be adapted
to local needs by means of its configuration files. All settings for the
VisDic application are stored in several XML files.
2.2.3.1 Global Configuration
The main configuration file (visdic.cfg) serves for global application
data storage such as the list of dictionaries, the list of views, fonts, colors
or query history. All information is stored in an XML structure.
The first-level subsections of the global configuration are:
colors In the COLOR section the user can define colors which are then
referenced by its name in dictionary configuration files. Each color
is enclosed in its name tag and consists of three hexadecimal values
separated by commas, representing consequently its red, green and
blue components. Each value can be in range 0x0000, 0xffff .
fonts The FONT section defines fonts which can be referenced by the
defined names from dictionary configurations. Each font definition
is enclosed in its name tag and its value correspond to the font
string description.
application settings The APPL section contains all global data that are
related to the application state. The most common settings that
can be found here are:
• DICT – path to a dictionary that is presented in the list offered
to the user.
• OPEN – ordinal number of a dictionary that should be opened
in one notebook tab.
31
2.2.3 2. NEW LANGUAGE RESOURCES AND TOOLS
• AUTOLOOKUP – definition of a synchronization link between two
notebook tabs.
• SIZE – size of the notebook tab in percentage of the main
window width.
• HIST – history of last queries that were entered by the user in
a specific notebook entry line.
A shortened example of a global configuration file is displayed in Figure
2.10.
2.2.3.2 Dictionary Specific Configuration
Each WordNet dictionary has its special configuration file (named dictionary.cfg),
which enables the linguist to set up most of the texts displayed
in the application as well as the content of notebook tabs specific to the
particular dictionary with respect to the XML structure of the entries.
The configuration contains attribute settings of the dictionary and
sections describing the layout of the dictionary views. The main attributes
in the dictionary configuration are:
• NAME – full name of the dictionary. This name is presented to the
user in various places in the application, e.g. on the top of the
dictionary notebook tab.
• SHORT NAME – short name of the dictionary.
• MAIN TAG – the default XML tag in the user queries (e.g. SYNSET.
SYNONYM.LITERAL).
• MAX QUERY – limit of the number of results of a query.
• MAX VIEW – limit of the number of characters displayed in the user
defined text view.
• CHARSET – name of a character set indicating the encoding of the
dictionary. This information is necessary for correct manipulation
with the dictionary in some systems.
The rest of the dictionary configuration file contains sections defining
the list of the available dictionary views and their content or the list of
duplicate checking actions in the application menu.
32
2. NEW LANGUAGE RESOURCES AND TOOLS 2.2.3
<?xml version="1.0"?>
<CONFIG>Visdic general configuration file
<COLOR>Colors definition
<BLACK>0x0000, 0x0000, 0x0000</BLACK>
<WHITE>0xffff, 0xffff, 0xffff</WHITE>
<RED>0xffff, 0x0000, 0x0000</RED>
...
</COLOR>
<FONT>Fonts definition</FONT>
<APPL>
<DICT>/nlp/wn/visdic/data/eng20/wneng20</DICT>
<DICT>/nlp/wn/visdic/data/cze/wncze</DICT>
...
<AUTOLOOKUP>v2-v1</AUTOLOOKUP>
<OPEN>1
<SIZE>43</SIZE>
<HIST>
<LINE>trench</LINE>
<LINE>house</LINE>
...
</HIST>
</OPEN>
<OPEN>5
<SIZE>57</SIZE>
<HIST>
<LINE>pes</LINE>
</HIST>
</OPEN>
</APPL>
</CONFIG>
Figure 2.10: The global configuration example (... stands for shortened
parts).
33
2.2.3 2. NEW LANGUAGE RESOURCES AND TOOLS
Visual Definitions The VISUAL section describes the way, how to display
dictionary entries. Definitions are enclosed in tags corresponding to
their names. VisDic uses primarily two special visual definitions. The
first is called VISDIC SHORT and it presents the entry in a short one-line
format (e.g. list of all entries matching the query or within a tree view).
The second visual definition, named VISDIC, describes the content of the
user defined text view, i.e. it presents the entry in a more descriptive way.
Each tag from the entity XML structure can be displayed in its own
way. The definition contains C-like string format specifications consisting
of a string in double quotation marks and other parameters. These
parameters have the following meaning:
• %c – color name (taken from visdic.cfg), changes the current
color.
• %f – font name (taken from visdic.cfg), changes the current font.
• %s – string, it can be @tag:name for the current tag name, or
@tag:value for the tag value.
• %i – includes the output of processing of subtags.
• %K+ – in the tree view, stop expanding the Full expansion view
under the current entry.
• %K – in the tree view, delete the current line from the tree.
The format string can include parts that are displayed only under a certain
condition. The available conditions are
• \\{^...\\} – display ... only if the tag is the first in the list.
• \\{$...\\} – display ... only if the tag is the last in the list.
• \\{*...\\} – display ... only if the tag is not the last in the list.
An example of usage of the conditional parts of the format string can be
a comma separated list of literals with their senses:
<SYNONYM>"%i"
<LITERAL>"%s:%i\\{*, \\}",@tag:value
<SENSE>"%s",@tag:value</SENSE>
</LITERAL>
</SYNONYM>
34
2. NEW LANGUAGE RESOURCES AND TOOLS 2.2.3
The visual definition of each XML tag can contain a test for the value
of the tag in the form <TAG>="value":"result"</TAG>. For instance,
various type of WordNet relations between synsets can be transcribed in
colored one-letter acronyms like this
<ILR>"%i"
<TYPE>="hypernym":"%cH",BLACK</TYPE>
<TYPE>="holo_member":"%cM",BLUE</TYPE>
<TYPE>="derived":"%cD",DARK_GREEN</TYPE>
<TYPE>"%c[%s]",RED,@tag:value</TYPE>
</ILR>
A special tag named DEFAULT stands for any tag. It is used for tags that
do not have their own definitions.
Views The VIEW section specifies the design of notebook tabs. Each
tab is described by one LIST subsection. Each tab has its own name in
the NAME tag and its own type in the TYPE tag. According to the type,
the LIST subsection can include other specifications of the tab content:
• XML view has no other options. It just displays the XML structure
of a dictionary entry.
• USER type has DEF tag referencing the name of a visual definition
of the user defined text view.
• TREE type contains two tags specifying parent and child link tags
in the dictionary and the DEF tag for the visual definition used in
the presented tree-like ordering of entries.
• EDIT type describes the form fields for editing one dictionary entry.
The subsection contains ITEM or BUTTON tags. Items refer to XML
tags in TAG, each has its own head label in HEAD and its own item
type in TYPE. The appearance of the form field is specified in the
EDIT tag. It can be a single line entry (ENTRY), a multi line entry
(TEXT) or a checkbox (CHECKBOX). All form fields used for editing
the link or reverse link tags12
will be displayed as combo boxes with
an arrow on the right side of the box, which allows the user to
12see the field type R in the Section 2.2.3.3
35
2.2.3 2. NEW LANGUAGE RESOURCES AND TOOLS
navigate to the referred entry. All form fields that represent a tag
which can occur more than once are supplemented by two buttons
and . These buttons are used for adding another instance or
removing the current instance of the tag.
The BUTTON tags define buttons that run one of the storage actions.
Each button has its label specified in the TEXT tag and its type in
the TYPE tag. The type can be either NEW for creating the new
entry, DELETE for deleting the current entry or UPDATE for saving
the edited entry content.
• WORD type view presents a list of all words from the dictionary that
can be found among values of the given tag.
• ENTR type view is a list of entries that meet a condition given by
the user query in the QUERY tag.
Main Menu Actions The MENU section describes a list of dictionaryspecific
actions which can be added to the VisDic main menu. All these
actions will be appended to the Dictionary submenu.
An example of the actions that can be specified in the MENU are the
duplicate checking actions. These actions are looking for duplicate values
within the dictionary, either among entries or within a single entry. The
action definition is enclosed in the DUPL tag. The TYPE tag chooses the
kind of comparison – ENTR for comparing entries or ITEM for comparing
items within the range of one entry. The NAME contains a name of the
action, which will be displayed in the menu. The TAGS tag enlists all tags
that are included in the duplicate checking, more tags are separated with
the vertical bar sign ‘|’. If a tag begins with a dot ‘.’, then the tag is
considered as a subtag of the previous tag.
Examples of the duplicate checking actions are:
• searching for all entries (synsets in WordNet) having the same
SYNSET.ILI value
<DUPL>
<TYPE>ENTR</TYPE>
<NAME>Check duplicate ILI numbers</NAME>
<TAGS>SYNSET.ILI</TAGS>
</DUPL>
36
2. NEW LANGUAGE RESOURCES AND TOOLS 2.2.3
• identification of all pairs literal:sense in WordNet stored in more
than one synset. Here, the .SENSE tag corresponds to SYNSET.
SYNONYM.LITERAL.SENSE subtag of the SYNSET.SYNONYM.LITERAL
tag
<DUPL>
<TYPE>ENTR</TYPE>
<NAME>Check duplicate literals & senses</NAME>
<TAGS>SYNSET.SYNONYM.LITERAL|.SENSE</TAGS>
</DUPL>
• finding all literals in WordNet that occur more than once in one
entry (synset)
<DUPL>
<TYPE>ITEM</TYPE>
<NAME>Check duplicate synset literals</NAME>
<TAGS>SYNSET.SYNONYM.LITERAL</TAGS>
</DUPL>
2.2.3.3 Dictionary Definition
Each dictionary has, besides its configuration file, an associated definition
file named dictionary.def. This file describes the XML structure of the
dictionary. The structure of the definition file contains features that are
specific to the WordNet-like XML dictionaries.
The definition file format is a plain text with each row corresponding
to one XML tag. The line format is
level tag min max type args
where the corresponding fields contain
• level – the tag level (0 for the top level).
• tag – the tag name.
• min – minimal number of occurrences of the tag within its supertag.
• max – maximum number of occurrences of the tag within its supertag
(−1 means infinite number).
37
2.2.3 2. NEW LANGUAGE RESOURCES AND TOOLS
0 SYNSET 1 1 N
1 ID 1 1 K
1 POS 1 1 N
1 SYNONYM 1 1 N
2 LITERAL 1 -1 N
3 SENSE 1 1 I @maxbyparent+1
3 LNOTE 0 1 N
1 ILR 0 -1 L
2 TYPE 1 1 N
1 RILR 0 -1 R SYNSET.ILR
1 BCS 0 1 N
1 DEF 0 1 N
1 USAGE 0 -1 N
1 SNOTE 0 -1 N
1 STAMP 0 1 N
Figure 2.11: An example of a dictionary definition file.
• type – the kind of the tag. It can be one of
– N – normal text entry.
– I – integer number entry. In the args column a function for
the default value can be stated.
– K – key value uniquely identifying the entry. Such key is used
by the following L, R, and E kinds of tags.
– L – link to another synset, it represents a semantic relation.
– R – similar to L. It is defined as reversed link specified in the
args column.
– E – contains an external information stored in another dictionary.
The name of the external tag and the path to the
dictionary are contained in the args columns. The path is absolute
or relative to the VisDic initial directory, not relative
to the dictionary path.
• args – extra arguments for some kinds of tags.
An example of a dictionary definition file can be found in the Figure 2.11.
38
2. NEW LANGUAGE RESOURCES AND TOOLS 2.3
2.3 DEBVisDic and other DEB Platform Ap-
plications
VisDic, during its rather short history, has already proved its suitability
for lexical database creation. The main power of VisDic manifests
itself especially in development of highly interlinked databases such as
WordNet. Its unique features have assured VisDic the leading role in
many WordNet editing projects.
In comparison with previous WordNet tools, VisDic exploits XML
data format thus making the WordNet-like databases more standard and
exchangeable. Not only that, thanks to the XML data format used and
to its dictionary specific configurability VisDic can serve for developing
various types of dictionaries, i.e. monolingual, translational, thesauri and
multilingually linked WordNet-like databases. The experience with the
VisDic tool during Balkanet project has been positive [HS04] and it was
used as the main tool with which all Balkanet WordNets were developed.
VisDic, however, has its disadvantages, particularly it is not based on
the client/server architecture and it does not allow to associate various
attributes with literals and handle the links between them. It can work
with links only between synsets which is a limiting feature for enriching
WordNets with various sorts of information, e.g. in Czech with word
derivation relations existing within one part of speech as well as across
them.
The experience with VisDic has led us to more systematic research
into the usage of XML data formats within the field of the computational
lexicography. In parallel, we also pay attention to the relations between
WordNets and Semantic Web. This interest gives us a strong motivation
for studying the properties of the XML data formats and tools for working
with them.
Thus we set as our task to design and implement a more universal
dictionary writing system that could be exploited in various lexicographic
applications to build large lexical databases. The system has been called
Dictionary Editor and Browser (further DEB) and in its current version
(named DEB ii) is used in several larger lexicographic projects that are
described further. The design of DEB allows us to use it advantageously
also for building WordNet-like databases.
39
2.3.1 2. NEW LANGUAGE RESOURCES AND TOOLS
Figure 2.12: The schema of the DEB II platform architecture
2.3.1 The Features of the Platform for Lexicographers’
Tools
The acronym DEB ii denotes a platform or framework for building (especially)
dictionary writing applications. It is based on client/server architecture,
thus the application falls into two parts (see the schema on the
Figure 2.12). The server includes majority of the required functions, each
client part on the other hand serves as a user graphical interface which
transfers user’s requirements to the server that returns the demanded
data. The server part is built from small parts, called servlets, which
allow a modular composition of all services.
The clients communicate with servlets using HTTP requests in a manner
similar to recently popular concept in web development called AJAX
(Asynchronous JavaScript and XML [RM98]). The data are transported
(using plain HTTP) in the RDF, generic XML or plain-text formats or
they are marshalled using JSON (JavaScript Object Notation [Cro06])
data structure encapsulation.
The actual data storage backend on the server side is provided by Oracle
Berkeley DB XML [DBX07, CRZ03, BS05], which is a native XML
database providing XPath and XQuery access into a set of document containers.
The metadata are stored in widely-used Berkeley DB embedded
40
2. NEW LANGUAGE RESOURCES AND TOOLS 2.3.1
database which runs on many systems and devices ranging from Linux
and Windows operating systems to mobile phones. Oracle Berkeley DB
XML comes in form of a C++ library with interfaces to many scripting
languages.
Since the client applications are mostly oriented to the graphical user
interfaces (GUI), we have decided to adopt the concepts of the Mozilla Development
Platform [O+
02]. Firefox Web browser is one of the many applications
created using this platform. Other applications include Mozilla
Thunderbird mail client, Netscape Web browser, Komodo integrated development
environment or Nvu web page editor.
The Mozilla Cross Platform Engine provides a clear separation between
application logic and definition, presentation and language-specific
texts. The application design is simple and allows the possibility of concurrent
work of different team members which leads to significant time
savings.
Mozilla platform is open source free software which ensures that it
will stay free and its development will continue. Every new major version
adds more features and possibilities. Also, thanks to open source design,
there is a large number of free extensions of existing applications or the
platform itself. Mozilla developers pay much attention to security and
any reported bugs are promptly fixed.
Applications built on the Mozilla platform are working within many
operating systems, actually any OS on which Mozilla runs (i.e. officially
Windows, Linux, and Mac OS X, unofficially many others).
The platform also provides easy way (both for developers and users)
for application installation and update.
The main “programming language” used for the GUI design of the
DEB clients is called XUL (XML User-interface Language, pronounced
“zool”). XUL is a user interface description language based on XML.
It allows relatively simple creation of cross platform applications with
possibility of easy customization of design, texts and localization. XUL
itself is aimed mostly on creation of user interfaces, e.g. windows, buttons
or toolbars, but it incorporates wide range of standardized technologies:
• Cascading Style Sheets (CSS) for describing the graphic appearance
of the application,
• JavaScript as a programming language for simple application logic,
• Document Object Model (DOM), XSLT and XPath to work with
HTML and XML documents,
41
2.3.1 2. NEW LANGUAGE RESOURCES AND TOOLS
• DTD for easy localization,
• RDF as data source.
2.3.1.1 The DEB Server Side
The server side of DEB is implemented in the programming language
called Ruby [TH01, RUB07]. Ruby (originating in Japan) is an objectoriented,
interpreted programming language with weak type checking.
The DEB server uses also various additional libraries, both pure Ruby
and interfaces to C/C++ libraries:
• REXML (XML processing) and the WEBRick HTTP and SOAP4R
SOAP servers (client–server communication). These modules are
pure Ruby.
• Oracle Berkeley DB XML API (storage backend).
• ICU (International Component for Unicode) by IBM – language
dependent character manipulation, sorting and formating [ICU07],
libxslt and libxml2 (XSLT and additional XML processing) from
the GNOME project. We actively participate in the development
of both ICU and libxml2/libxslt bindings to the Ruby programming
language.
• the GRASS geographical information system13
(GIS) interface used
in several client applications for displaying the geographical linkage
of the linguistic data.
• SQL interface (connection to classical relational databases, used
e.g. for the database of geographical data).
The DEB server suite runs on the Linux operating system, currently it is
tested with Ubuntu Dapper on Intel x86 and AMD64 architectures, but
it should generally run on any recent UNIX-based system (including Mac
OS X).
Current DEB server modules (i.e. the servlets) include:
• generic document servlet – serves data from a DB XML container,
supports querying the database, fetching individual documents and
storage of documents or XSLT transformation of the output;
13see [NM04, GRA07]
42
2. NEW LANGUAGE RESOURCES AND TOOLS 2.3.2
• SQL servlet – provides interface to relational data in PostgreSQL
(or other SQL) database;
• various specific servlets based on generic document servlet – provide
additional function over XML data stored in the DB XML
container;
• GRASS servlet – provides interface to the GRASS GIS, it is used
for map generation;
2.3.2 Assets of the DEB Platform
The DEB platform is based on client-server architecture, which brings
along a lot of benefits. All the data are stored on the server and a
considerable part of the functionality is also implemented on the server,
while the client applications can be very lightweight.
This approach provides adequate basis for team cooperation; data
modifications are immediately seen by all the users. The server part also
provides authentication and authorization mechanisms.
The server can offer different interfaces using the same data structure
and these interfaces can be reused by many client applications. For example,
several client applications use the same interface to query XML
based dictionaries (with different underlying structure).
Although the clients are usually created using the Mozilla platform,
the client software can be implemented in any way – it may be coded in
any programming language or may even look only as a simple web page.
One of the main benefits of developing a dictionary writing system on
the DEB platform is the homogeneity of the data structure and presentation.
If the application administrator commits a change in the data presentation,
this change will automatically appear in each client software.
And of course, any data flaws discovered can be instantly corrected, there
is no need to change the client software or provide new data files to each
client.
The data sources can be implemented with different structures, that
the server transforms seamlessly to a homogeneous form, which is then
provided to client applications.
Of course, a drawback of the client-server architecture is that an operating
server is necessary for a fully functional application. However, in
special situations, the server can be installed within a local environment,
43
2.3.3 2. NEW LANGUAGE RESOURCES AND TOOLS
or for the possibility of simple off-line WordNet editing, the client may
work in a degraded manner without the instant connection to the server.
2.3.3 The DEB Administration Interface
Initially, the DEB server was developed with just command-line management
of dictionaries and administration of user passwords for authentication.
The configuration was realized by structured text files and data
processing scripted programs.
After the client applications have spread to more users world-wide and
have been used, e.g., in several national WordNet projects (Dutch, Polish,
Hungarian, Slovenian or Afrikaans WordNets), a more sophisticated
administration interface for the DEB users and dictionaries was created
by Adam Rambousek in the MU NLP Centre under the supervision of
Aleš Horák. The interface was gradually transformed into a general and
complex dictionary management application for the whole DEB server.
2.3.3.1 Overall Design Goals
The DEB server packages are currently being deployed on several servers
in different organizations and often more than one user need to administer
a single DEB server without having a direct server access. Thus, the
administration interface must be accessible remotely and without any
special tools. The best choice for this task is a web-based interface,
where the user needs just a web browser.
The interface should support easy administration of all the server
areas. Of course, the main area of a dictionary management server is
the dictionary management. Each dictionary is described with several
basic attributes, like its name and code, the filename of its storage in
the DB XML database, its dictionary type, the XML schema or indexed
elements or XSLT templates for output displaying. Also, some projects
may need extra specific settings – e.g. the DEBVisDic clients need to
store information about the inter-dictionary links. After the dictionary
is set up, the interface has to support import and export of XML data
into and from the DB XML format.
When the administrator sets up the server dictionaries, these can be
grouped to “services.” A service is one individual part of the DEB server,
usually used for one particular project. For example, DEBVisDic or
44
2. NEW LANGUAGE RESOURCES AND TOOLS 2.3.3
DEBDict are separate services, but they share the same base libraries and
management database. Several services can access the same dictionaries,
each providing different view on the data.
The user accounts are shared between all the services. Thanks to the
database sharing between services, each user needs just one account for all
the services he or she may use. The administrator can restrict access to
selected services and for each service, more detailed access permissions
can be set for each dictionary (read-only, read-write, update, . . . , see
the Figures 2.13 and 2.14). The actual usage of the dictionary access
permissions depends completely on the service implementation. This
means, one service can ignore permissions at all and another service can
use complex access rights.
Apart from access rights, the user account management provides all
the needed functions – it allows to create, modify and delete user accounts.
Each user can log-in to the administration interface and change
his or her password. In case the user forgets a password, he or she can
ask for a new random password.
To ease the deployment of the DEB platform, we are experimenting
with automated creation of the client applications. Now, the server
is able to create straightforward applications based on the Relax NG
Schema [vdV03] of the dictionary, and we are aiming at automated creation
of client packages for new national WordNets.
Another very useful feature is uploading the client source files onto
the server using the web interface. This way, the administrator can easily
modify web page templates (XSLT) or other files without the need of
direct (FTP, SSH) access to the server.
The server administration interface is based on the same postulates
as the other DEB server dictionaries and modules. The Oracle Berkeley
DB XML database provides a storage backend for the administration
meta-data. The server-side scripts are developed in Ruby programming
language.
All the data about users, dictionaries, permissions and other control
data are stored in the DB XML database in the XML format. Each
dictionary module of the DEB server uses a common interface to access
data from this administration database.
The administration module provides several services – user authentication,
access rights control, entry locking and journaling of dictionary
changes.
45
2.3.3 2. NEW LANGUAGE RESOURCES AND TOOLS
Figure 2.13: User management showing how access rights modify the
dictionary list in DEBDict; list for selected user is on the left, list of all
dictionaries is on the right.
The administration interface is a web-based application where the
web pages are generated using an HTTP template which allows easy
design and content modification and then served to the users by a lightweight
web server – WEBrick [San04]. The users are authenticated using
standard HTTP authentication mechanism. The administration module
extends the standard interface for passwords stored in a file and loads
user’s login and password from the XML database. Each change in user
accounts or access rights is propagated to all DEB services in the real-
time.
2.3.3.2 The Dictionary Management
For each dictionary, the administrator has to define several attributes
(see the Figure 2.15). The minimal set of attributes contains a unique
dictionary code, a database filename and a dictionary class (the implementation
class in Ruby), the other attributes are more or less optional.
46
2. NEW LANGUAGE RESOURCES AND TOOLS 2.3.3
<user>
<login>adam</login>
<name>Adam Rambousek</name>
<email>xrambous@fi.muni.cz</email>
<org>Faculty of Informatics</org>
<addr>Botanicka 68a, Brno</addr>
<pass>3Ja8ivX12OB0U</pass>
<services><service code="debdict">
<dict code="scs" perm="r"/>
<dict code="scfis" perm="r"/>
<dict code="cia" perm="r"/>
<dict code="scfin" perm="r"/>
<dict code="diderot" perm="r"/>
</service></services>
</user>
Figure 2.14: XML entry for the user from the Figure 2.13.
Figure 2.15: Dictionary management showing basic information and indexed
elements for the Czech WordNet dictionary.
47
2.3.3 2. NEW LANGUAGE RESOURCES AND TOOLS
The meaning of the dictionary attributes is:
• the dictionary name is displayed to users by the client application.
• the definition of the XML entry root tag and its key element are
needed for XML import and for searching (in case, the application
does not have its own, more complex search method).
• indexes speed up search operations, so each element or attribute
that is used in user queries should be indexed.
• XSLT templates transform XML data to another form suitable for
presentation or machine processing.
Extra dictionary attributes are required for the WordNet dictionaries:
• each WordNet dictionary is linked to the client software by the
client package code.
• the WordNet Dictionaries can refer to each other using the specified
“equivalence tags.”
• in the next field, the administrator can list dictionaries that should
be reloaded after an edit action in the client (usually in another
dictionary).
• and the last option specifies related dictionaries – for example, several
national WordNets linked with ILI (Inter-Lingual Index). It is
possible to display the same entry in different languages or to copy
entries between languages.
2.3.3.3 Import and Export
The import function takes an XML file and stores the data into the
DB XML database. The XML file has to be uploaded to the server (it
is possible to upload it through web interface). All entries must share
the same root tag (specified in the dictionary management), entries with
different root tags are ignored. The administrator can choose if he or she
wants to delete all the entries from database before the import or just add
the new entries. The import utilizes two methods for XML reading. The
first method loads the whole XML file into memory and uses an XML
48
2. NEW LANGUAGE RESOURCES AND TOOLS 2.3.3
parser on the big document. This method is accurate, unfortunately
it has exponential time complexity, so it can take hours for large XML
files (over 10 MB). The second method uses regular expressions to read
entries one by one from the XML file and then each single entry is parsed.
Entries are stored in the database with value of the specified key tag as a
unique key. The administrator is informed about the import progress on
the web page – a number of processed entries, a total number of entries,
an estimated time till the end and last ten entry keys are displayed.
The administration module also supports export from database to
plain XML file, the output files may be compressed to save disk space.
The export also has an option to save the file in the form of a Ruby
language script that will setup the database and import initial data. This
is needed for the administration database itself. The output files are saved
in a specified directory on the server and the administrator is informed
about the export progress. Once the export ends, the administrator is
offered a link to download the file through the web interface. The same
function is used also for daily database backup.
2.3.3.4 Locking and Sequences of Identifiers
The administration interface offers entry locking management to other
DEB server modules. If multiple users can edit the database at the same
time (which is one of the basic advantages of the client-server architecture),
it is crucial to provide exclusive write locking of entries so that
two users are not able to edit the same entry at a time. Decisions about
entry locking depends on each application design:
1) when should an entry be locked and unlocked?
2) should only the edited entry be locked or should the locking affect
other entries too?
An application then sends the request to the administration module
which updates the lock database. The administration module provides
several functions – besides simple lock and unlock functions, it can tell
which user has locked a given entry, return the list of locks for selected
user and/or dictionary or group several locks together if they are related.
The administrator has access to the list of all locks and he or she can
also delete chosen locks if the application did not release them correctly.
Newly created entries should have a unique identifier. If the application
does not generate its own identifiers, the administration module can
49
2.3.4 2. NEW LANGUAGE RESOURCES AND TOOLS
provide such service. It is possible to set an identifier pattern for each
dictionary – this pattern looks like CZE-[id] and [id] will be replaced
with sequentially increased number. The administrator can also affect
the number used.
2.3.3.5 The Installation Packages
The administration interface supports automated creation of Firefox Extension
installation packages (XPI). If the administrator specifies a Relax
NG schema for the dictionary, it is possible to automatically transform
this schema to an application design description in the XUL description
language and the supporting code in JavaScript. The application created
in this way supports basic forms – single and multiple text fields,
select-boxes of specific values or relational links to other dictionaries. It
can serve as basis for custom modifications. Of course, the application is
able to connect to server, load data from server and save a modified entry
back. We are currently working on more complex support for creation of
new packages, mainly for the DEBVisDic client packages.
2.3.4 How To Make a Sample Dictionary
2.3.4.1 New Dictionary Definition
As a first step, the administrator needs to provide basic information about
the dictionary. The dictionary data can be loaded from an XML file or it
can be built from scratch. The administrator must specify an entry root
element, the XPath specification of a unique key, several indexes for fast
querying and an XML schema of the entry.
Let us create a demonstration dictionary from scratch, we will name
the root element entry and have the unique key identifier in the element
/entry/headword. The corresponding Relax NG schema is given in the
Figure 2.16.
This schema describes entry with one headword element, with pos
attribute, and one or more sense elements. Of course, Relax NG supports
description of much more complex XML structures.
50
2. NEW LANGUAGE RESOURCES AND TOOLS 2.3.4
<element name="entry">
<element name="headword">
<attribute name="pos"> <text/> </attribute>
<text/>
</element>
<oneOrMore>
<element name="sense"> <text/> </element>
</oneOrMore>
</element>
Figure 2.16: A part of the Relax NG dictionary schema.
Figure 2.17: An example client application generated by the DEB administration
interface according to the dictionary schema from the Figure
2.16.
2.3.4.2 Preparation of an Installation Package
The preparation of a new basic client application package requires selection
of a dictionary and running the package generation function. The
administration module checks the Relax NG schema and finds all elements
or attributes that contain the text child element. All such elements
and attributes are transformed to XUL textbox fields with the
respective name as a label describing the field. If an element can occur
multiple times in the entry (such as the sense tag in our example), buttons
for adding and removing the textbox are added to the application
form, too.
The created JavaScript supports loading and saving documents and
51
2.3.4 2. NEW LANGUAGE RESOURCES AND TOOLS
Figure 2.18: A web service automatically built following a dictionary
schema
also searching for documents. The application thus enables querying
each indexed field specified in the dictionary management interface. For
example, users can easily find all nouns.
All the created application files are then packaged into the Firefox
extension installation package (XPI). Users can download this package
for installation or individual files for editing. An example of the resulting
application is shown on the Figure 2.17.
For the new client, there are also two basic preview templates (in
XSLT) saved on the server side. One provides basic entry preview displaying
all the data and the second displays raw XML data.
For certain environments that either do not allow users to install new
software packages or where the deployment of the software would be too
time consuming, the DEB ii server is able to generate simple web-service
(see the Figure 2.18). The same as for XPI package generation, this function
uses the dictionary Relax NG schema and generates a XUL form for
52
2. NEW LANGUAGE RESOURCES AND TOOLS 2.3.4
a) <textbox id="entry.headword.@pos"/>
b) <menulist id="entry.headword.@pos">
<menupopup>
<menuitem label="noun"/>
<menuitem label="verb"/>
<menuitem label="adjective"/>
<menuitem label="adverb"/>
</menupopup>
</menulist>
Figure 2.19: Change of a textbox field to a drop-down list.
remote access. To work with the dictionary, a user needs a web browser
based on the Mozilla engine (Firefox, SeaMonkey, Netscape, Camino,
. . . ). All parts of the generated web-service are easily customizable via
XSLT templates.
2.3.4.3 Application Customization
Thanks to the design of applications based on the Mozilla development
platform, these applications are easily customizable.
Any change in the layout and design of the form is done by editing the
XUL (XML User-interface Language) files accompanied with standard
CSS stylesheets. The application logic (i.e. procedures implemented in
JavaScript) stays the same for a new layout. Combination of XUL and
CSS languages is very powerful and supports long list of features that are
commonly used in desktop applications. To give a simple example, we can
show how to change the Part-Of-Speech textbox field into a drop-down
list, see the Figure 2.19.
The application localization (translation of the user interface to another
language) is one of the core features of the Mozilla XPI packages.
As we can see in the example, the field labels contain XML element
names only (marked with &). This allows the application designer to
change them to general textual labels that are more informative to users.
The actual texts are stored in a DTD (Document Type Definition) file
as XML entity definitions, where they can be adjusted to any texts in
one place. For the sake of the above mentioned localization of the application,
it is possible to include several DTD files for different languages
53
2.3.5 2. NEW LANGUAGE RESOURCES AND TOOLS
application.xul: <label value="&entry.headword;"/>
en-US/application.dtd: <!ENTITY entry.headword "headword">
cs-CZ/application.dtd: <!ENTITY entry.headword "heslo">
Figure 2.20: A field label and the respective entity in the localized DTD
files.
into the installation package and (automatically) switch between them
(see the Figure 2.20).
After all the application source files are modified to meet the designer’s
requirements, he or she can upload them using the administration
interface and let it build a new version of the installation package.
The application designer can also supplement the dictionary editor
with more preview templates or modify the existing ones for different
data presentation. When adding a new template, the template name
must be added to the dictionary description in the database management
interface. The modified templates are again uploaded to the server via
the administration interface.
2.3.5 Usage Variability – The Users’ Interfaces
The DEB clients are written in XUL and JavaScript and integrate with
Mozilla Firefox web browser. This allows the developers to use both
Mozilla’s user interface engine and its HTML/XHTML rendering engine
as well as built-in components for interaction with filesystem on client
computers, XPath interpreter, RDF processor etc.
Due to the feature-rich client architecture the developers may decide
whether certain operations should be done on the server or on client
parts – e.g. XSLT transformation can be done on both sides.
The particular DEB clients that are currently being implemented
within the DEB platform include:
• DEBVisDic – complete new version of the successful WordNet semantic
network editor and browser VisDic, see the Section 2.3.5.2.
• DEBDict – a general dictionary browser.
54
2. NEW LANGUAGE RESOURCES AND TOOLS 2.3.5
Figure 2.21: The DEBDict common interface to several dictionaries with
different structures.
• Czech Onomastic Dictionary – newly prepared dictionary of Czech
proper names and their origins
• PRALED – implemented in cooperation with the Institute of Czech
Language, Czech Academy of Sciences in Prague. The task with
PRALED is to build a new Czech Lexical Database. See the Section
2.3.5.4 for more information.
• DEB CPA – tool used for several distant teams to work on a new
resource of Corpus Pattern Analysis. It is further described in the
Section 2.3.5.3.
• DEB TEDI – the main tool for preparation of a new terminological
dictionary of Czech art terms. This work is a joint project
of the Faculty of Fine Arts, Brno University of Technology and
Masaryk University. The aim of the project is to build a terminological
database consisting of about 5 000 dictionary entries which are
classified into categories and supplemented with term definitions,
translations info English, German and French, and with Czech usage
examples. The resulting dictionary will be offered as a publicly
55
2.3.5 2. NEW LANGUAGE RESOURCES AND TOOLS
available application directed especially to fine arts students.
• Cornetto – several client tools working over a complex database
schema of three interconnected dictionaries of Dutch. Prepared in
cooperation with the University of Amsterdam. Details are presented
in the Section 2.3.5.5.
• VisualBrowser – the DEB ii client-server architecture allows an easy
connection of other existing applications to the DEB server. An example
of such application is a direct interface to the VisualBrowser
tool [Nev05] that now displays the graphical representation of relations
between elements stored in various DEB server databases.
2.3.5.1 DEBDict
The dictionary browser DEBDict originally started as a demonstration
application for the DEB platform and it gradually evolved into an indispensable
resource for several hundreds of linguists in several countries.
The DEBDict functionalities include:
• multilingual user interface (English and Czech, others can be easily
added)
• queries to several XML dictionaries (of different underlying structure)
with the result passed through an XSLT transformation
• integration with Czech morphological analyzer
• connection to an external website
• connection to the GRASS geographical information system (display
of geographical links directly on their positions within a cartographic
map) or any similar application
The version of DEBDict that is currently running on the MU NLP Centre
server provides a common interface to 7 dictionaries (see the Figure 2.21):
• the Dictionary of Literary Czech Language (SSJČ [P+
02], 180 000
entries)
• the Reference Dictionary of Czech Language (PSJČ [Hav57], with
200 000 entries)
• the Dictionary of Literary Czech (SSČ [F+
95], 49 000 entries)
• the Dictionary of foreign words (SCS [KP+
99], 46 000 entries)
• the Dictionary of Czech Synonyms (thesaurus [PJ94], 23 000 en-
tries)
56
2. NEW LANGUAGE RESOURCES AND TOOLS 2.3.5
Figure 2.22: The DEBVisDic main interface
• two parts of the Dictionary of Czech Phrasal Words and Idioms
(shortly SČFI [Č+
83], 4 000 verbal entries and 10 000 non-verbal
entries)
• the Diderot encyclopedia (90 000 entries)
As an addition, DEBDict features an interconnection to several web systems
(such as Google or Answers.com) and the GRASS geographical
system with the list of the Czech towns and cities.
57
2.3.5 2. NEW LANGUAGE RESOURCES AND TOOLS
Figure 2.23: The DEB CPA tool.
2.3.5.2 DEBVisDic
DEBVisDic was one of the first applications built over the DEB ii platform
– it was designed as a completely new client-server tool for WordNet
browsing and editing that should serve as a successor of the VisDic program
presented in the Section 2.2.
DEBVisDic uses new versatile interface (see the Figure 2.22 that allows
the user to arrange the work without any limitations. Of course,
DEBVisDic contains all the main features that were present in VisDic:
• multiple views of multiple WordNets
• freely defined text views
• synset editing
• hypero-hyponymic tree
• query result lists
• plain XML view of a synset
• synchronization
• inter-dictionary linking
58
2. NEW LANGUAGE RESOURCES AND TOOLS 2.3.5
• tree browsing
• consistency checks
• journaling
• user configuration
With the help of the DEB platform reusability, DEBVisDic will be supplemented
with many new features that are currently accessible only as
separate tools or resources. This functionality includes:
• connection to a morphological analyzer (for languages, where it is
available)
• connection to language corpora, including Word Sketches statistics
• access to any electronic dictionaries stored within the DEB server
• searching for literals within encyclopedic web sites
• and many others
Currently, DEBVisDic is also used for preparation of new Polish, Hungarian,
Slovenian, Dutch (within the Cornetto project) and Afrikaans
WordNets and it is proposed as the main tool for the prepared Global
WordNet Grid.
2.3.5.3 DEB CPA Editor and Browser
Corpus Pattern Analysis (CPA, [Han04]) is a new technique for mapping
meaning to words in text. No attempt is made in CPA to identify
the meaning of a verb or noun directly, as a word in isolation. Instead,
meanings are associated with prototypical sentence contexts. Concordance
lines are grouped into semantically motivated syntagmatic patterns.
Associating a “meaning” with each pattern is a secondary step,
carried out in close coordination with the assignment of concordance lines
to patterns.
CPA editing tool (see the Figure 2.23) displays the list of verb entries,
along with the information who and when updated the entry. Each entry
consists of several patterns (the number of patterns is not limited) and
it is possible to freely modify their order and content. The main part of
the tool, the pattern editing window, allows to enter and modify all the
information about one pattern. The form is very versatile, e.g. it allows to
add any number of subject/object alternations. The tool is connected to
an on-line resource – it is possible to look up subject and object semantic
type in Brandeis Semantic Ontology [PHL+
06] which is hosted on a web
59
2.3.5 2. NEW LANGUAGE RESOURCES AND TOOLS
Figure 2.24: The PRALED user interface
server at Brandeis University. Examples documenting the pattern are
taken from BNC using a modified version of Bonito2 corpus manager
that is integrated to the DEB CPA tool.
2.3.5.4 The PRALED Lexicographic Station
This client is designed for the development of the Czech Lexical Database
(CLD, denoted also as LEXIKON 21 [RK07]) and it serves as a main
tool in preparation of the new comprehensive and exhaustive database
of lexicographic information for the Czech language. The user’s part of
the PRALED tool is presently under the development in the Institute of
Czech Language (ICL), Czech Academy of Sciences, Prague.
The PRALED system offers the following functionality:
• queries to several XML dictionaries (of different underlying structures),
particularly to all relevant Czech dictionaries, i.e. SSJČ,
60
2. NEW LANGUAGE RESOURCES AND TOOLS 2.3.5
SSČ, SCS, SČFI and DIDEROT (see [P+
02, F+
95, KP+
99, Č+
83]),
• editing existing or writing new dictionary entries. A lexicographer
can use a set of forms which define the structure of the entry and
fill in all relevant fields (see the Figure 2.24) which presently are:
– pronunciation and spelling
– morphological properties (POS, the respective grammatical
categories)
– description of the meaning (entry definition)
– word formation nest (subnet)
– syntactic properties (most often valencies)
– stylistic, domain and regional features
– semantic relations to other entries (cross-references)
– etymological information
– integration with Czech morphological analyzer
– connection to an external website (Google, or Answers.com)
– remarks and additional comments
– integration with the corpus manager Bonito2 and Word Sketch
Engine [KRST04], which allows a lexicographer to obtain the
sorted individual word contexts including frequencies and statistical
distribution parameters (salience).
2.3.5.5 Cornetto
The Cornetto project is a Dutch national project funded in the STEVIN
framework. The goal of this project is to build a lexical semantic database
for Dutch, covering 40K entries, including the most generic and central
part of the language. Cornetto will combine the structures of both
the Princeton WordNet and FrameNet for English [FBS04], by combining
and aligning two existing semantic resources for Dutch: the Dutch
WordNet [Vos98] and the Referentie Bestand Nederlands [MMdM99].
The Dutch WordNet (DWN) is similar to the Princeton WordNet for
English, and the Referentie Bestand (RBN) includes frame-like information
as in FrameNet plus additional information on the combinatoric
behaviour of words in a particular meaning. The combination of the two
lexical resources will result in a much richer relational database that may
improve natural language processing (NLP) technologies, such as word
sense-disambiguation and language-generation systems. In addition to
61
2.3.5 2. NEW LANGUAGE RESOURCES AND TOOLS
Figure 2.25: Data collections in the Cornetto database (the schema was
prepared by Piek Vossen).
merging the WordNet and FrameNet-like information, the database is
also mapped to a formal ontology to provide a more solid semantic back-
bone.
The prepared Cornetto database (CDB) consists of three main data
collections:
1. collection of Lexical Units, mainly derived from the RBN
2. collection of Synsets, mainly derived from DWN
3. collection of Terms and axioms, mainly derived from SUMO [NP01]
and MILO [NT04]
In addition to the three data collections, a separate table of so-called
Cornetto Identifiers (CIDs) is provided. These identifiers contain the
relations between the lexical units and the synsets in the CDB but also
to the original word senses and synsets in the RBN and DWN. The
Figure 2.25 shows an overview of the different data structures and their
relations.
Since one of the basic parts of the Cornetto database is the Dutch
WordNet, we have decided to use DEBVisDic as the core for Cornetto
62
2. NEW LANGUAGE RESOURCES AND TOOLS 2.3.5
client software. We have developed four new modules, described in more
details below. All the databases are linked together and also to external
resources (Princeton English WordNet and SUMO ontology), thus every
possible user action had to be very carefully analyzed and described.
Cornetto Lexical Units The Cornetto foundation is formed by Lexical
Units, so let us describe their client package first. Each entry contains
complex information about morphology, syntax, semantics and pragmatics,
and also lots of examples with complex substructure. Thus one of
the important tasks was to design a preview to display everything needed
by the lexicographers without the necessity to scroll a lot. The examples
were moved to separate tab and only their short resumé stayed on the
main preview tab.
Lexical units also contain semantic information from RBN that cannot
be published freely because of licensing issues. Thus DEBVisDic here
needs to differentiate the preview content based on the actual user’s access
rights.
The same ergonomic problem had to be resolved in the edit form. The
whole form is divided to smaller groups of related fields (e.g. morphology)
and it is possible to hide or display each group separately. By default,
only the most important parts are displayed and the rest is hidden.
Another new feature developed for Cornetto is the option to split the
edited entry. Basically, this function copies all content of edited entry to
a new one. This way, users may easily create two lexical units that differ
only in some selected details.
Because of the links between all the data collections, every change in
lexical units has to be propagated to Cornetto Synsets and Identifiers.
For example, when deleting a lexical unit, the corresponding synonym
must be automatically removed from the synset dictionary.
Cornetto Synsets Synsets are even more complex than lexical units,
because they contain lots of links to different sources – links to lexical
units, relations to other synsets, equivalence links to Princeton English
WordNet, and links to the ontology.
Again, designing the user-friendly preview containing all the information
was very important. Even here, we had to split the preview to
two tabs – the first with the synonyms, domains, ontology, definition and
63
2.3.5 2. NEW LANGUAGE RESOURCES AND TOOLS
Figure 2.26: Cornetto Synsets window, showing a preview and a hyperonymy
tree
short representation of internal relations, and the second with full information
on each relation (both internal and external to English WordNet).
Each link in the preview is clickable and displays the selected entry in
the corresponding dictionary window (for example, clicking on a synonym
opens a lexical unit preview in the lexical unit window, see the
Figure 2.26).
The synset window offers also a tree view representing a hypernym/hyponym
tree. Since the hypero/hyponymic hierarchy in WordNet forms
not a simple tree but a directed graph, another tab provides the reversed
tree displaying links in the opposite direction (this concept was
introduced in the VisDic WordNet editor). The tree view also contains
information about each subtree’s significance expressed by the number of
direct hyponyms and the number of all the descendant synsets.
The synset edit form looks similar to the form in the lexical units
window, with less important parts hidden by default. When adding or
editing links, users may use the same queries as in dictionaries to find
the right entry.
Cornetto Identifiers The lexical units and synsets are linked together
using the Cornetto Identifiers (CID). For each lexical unit, the automatic
aligning software has produced several mappings to different synsets (with
different score values). At the very beginning, the most probable one was
64
2. NEW LANGUAGE RESOURCES AND TOOLS 2.3.5
Figure 2.27: Cornetto Identifiers window, showing the edit form with
several alternate mappings
marked as the “selected” mapping.
In the course of work, users have several ways for confirming the
automatic choice, choosing from other offered mapping, or creating an
entirely new link (see the Figure 2.27). For example, a user can remove
the incorrect synonym from a synset and the corresponding mapping will
be marked as unselected in CID. Another option is to select one of the
alternate mappings in the Cornetto Identifiers edit form. Of course, this
action leads to an automatic update of synonyms.
The most convenient way to confirm or create links is to use Map
current LU to current Synset function. This action can be run from any
Cornetto client package, either by a keyboard shortcut or by clicking on
the button. All the required changes are checked and carried out on the
server, so the client software does not need to worry about the actual
actions necessary to link the lexical unit and the synset.
65
2.4 2. NEW LANGUAGE RESOURCES AND TOOLS
Cornetto Ontology The Cornetto Ontology is based on SUMO and
so is the client package. The ontology is used in synsets, as can be seen
in the Figure 2.26. The synset preview shows a list of ontology relations
triplets – relation type, variable and variable or ontology term.
Clicking on the ontology term opens the term preview. A user can
also browse the tree representing the ontology structure.
2.4 Future Work on the Language Resources
and NLP Tools
We have displayed the details of the VerbaLex verb valency frames
dictionary and described the history of the design and implementation of
its complex valency frames format that was needed for encapsulation of
new semantic roles and links to the Czech WordNet entries. We expect
the version VerbaLex 2.0 to be available by the end of 2007 with the
number of 10 500 verbs fully linked to the Czech WordNet.
The nearest development of the VerbaLex dictionary includes linking
of the newly added verb synsets to the Czech and Princeton WordNets.
After enhancing the verb frame extraction algorithm in the syntactic
parser synt, we are going to perform large scale checking and extending
the lexicon by means of corpus text analysis.
We have described the DEB implementation platform and the main
features of applications built over this framework. The DEB platform advantageously
uses the client/server architecture and offers several different
clients allowing to perform various lexicographic tasks. The relevant
features of the DEB platform are high modularity and configurability.
Thanks to them, the DEB platform represents a versatile base, on which
the individual and powerful dictionary writing tools (clients) can be im-
plemented.
The presented off-line WordNet editor and browser, VisDic, during its
not so long history, proved its usefulness and contributed to the WordNetlike
databases creation especially within the Balkanet project. We have
shown that DEBVisDic as its successor retains its functionality and adds
new functions that will allow the lexicographers and researchers to create
new high quality lexical resources without which further progress in the
NLP field can hardly take place.
66
2. NEW LANGUAGE RESOURCES AND TOOLS 2.4
The development of DEBVisDic is also related to current Semantic
Web projects, in particular we will use the tool for building ontologies
covering various domains (one of the candidates is oncology). In this
connection, the above mentioned VisualBrowser [Nev05] is a suitable tool
for natural presentation of a semantic network.
Even though the DEB platform is developed as open source and free
platform, we believe that it offers interesting features for dictionary writing
systems and that it can speed up the development in this area. With
the new common administration module in the DEB development platform
that is shared by all kinds of DEB client applications, the DEB
platform provides a invaluable basis for new dictionary writing applications
for all purposes and all types of dictionaries. The applicability of
the platform is best justified with the implemented clients described in
this chapter and with the 250 registered users form 14 domains/countries
that currently access those applications at the Masaryk University DEB
server and with 6 other server installations in Prague, Amsterdam-UvA,
Amsterdam-VU, Poznan, Johannesburg and Budapest.
In the Cornetto project clients, we have presented how a combination
of automatic scored strategies with the human lexicographic work can
be used for merging large databases of previous dictionaries to obtain a
new qualitative language resource with complex morphological, syntactic
and semantic information. The presented project tools are, however, not
a single purpose programs but they fit in the general framework of the
DEB platform used for developing other publicly available language data
tools.
With the creation of the DEB ii platform and the individual clients
PRALED, DEBDict, or DEBVisDic lexicographers obtain new tools,
which can offer rapid development of necessary lexical resources, i.e. the
Czech Lexical Database in particular and a new dictionary of Czech later.
There have been no lexicographic station tools available for Czech users
so far. Our final aim is to equip the DEB ii platform with enough versatility
for all needs, and to make it easily portable to different system
installations.
67
2.4 2. NEW LANGUAGE RESOURCES AND TOOLS
68
Chapter 3
synt – Czech Syntax
Analyzer
The main tasks of automatic natural language (NL) processing are based
on a correct syntactic analysis of a NL sentence. While the quality of parsing
of analytical languages (English, German, . . . ) has already achieved
nearly satisfactory results [BBHS+
04, Rad93], the analysis of free word
order languages still faces many problems either in the form of huge
number of rules or output trees or it offers lower precision or coverage on
corpus texts [HS02, JG02, Hof95].
The syntactic analysis of running texts plays a crucial role in many
areas of advanced written and spoken text processing ranging from grammar
checking, machine translation or phrase identification to knowledge
mining and ontology acquisition. The problem of syntactic parsing is
often reduced to shallow syntactic analysis, e.g. [SL00], which is sufficient
for many applications where the speed of the processing is more
important than obtaining an exact and deep syntactic representation of
sentence. On the other hand, when the final aim is a thorough meaning
representation of the input sentence, a complete parsing is inevitable.
This is also the case of the synt system described further, which is being
implemented as a part of the Normal translation algorithm of natural
language sentences to constructions of Transparent intensional logic.1
1see [Hor02] and the Chapter 4
69
3. SYNT – CZECH SYNTAX ANALYZER
In the NLP Centre at Masaryk University, syntax parsing of the Czech
language as a representative of free word order languages forms one of the
main stream research tasks since the establishment of the Centre. The
most advanced system out of several implemented analysers is a parser
called synt [HK05a]. Recently, synt features a developed meta-grammar
of Czech and a fast parsing mechanism which offers a coverage of more
than 92 percent of Czech sentences2
while keeping the analysis time on
the average of 0.07s/sentence. The current development of the parser is
aimed at exploring several Best Analysis Selection methods that allow to
identify the preferred derivation tree in the output – due to the grammar
ambiguity the number of possible outputs (derivation trees) in synt is
often very high (in some cases hundreds of thousands and more). For
many reasons, this number is unacceptable for further processing and so
methods aimed at reducing it (while keeping the precision of the output)
are being developed as well as methods that derive structural information
directly from the packed shared forest obtained from synt.
The average number of parsing trees per input sentence strongly depends
on the background grammar and thence on the language. There
are natural language grammars producing one parsing tree but also highly
ambiguous grammar systems producing enormous number of results. Ambiguity
on all levels of representation is an inherent property of natural
languages and it also forms a central problem of natural language parsing.
A consequence of the natural language ambiguity is a high number
of possible outputs of a parser that are represented by labeled trees.
A grammar extracted from the Penn Treebank and tested on a set of
sentences randomly generated from a probabilistic version of the grammar
has in average 7.2 × 1027
parses per sentence according to Moore’s
work [Moo00a].
A traditional solution of these problems is presented by probabilistic
parsing techniques [BN00] aiming at finding the most probable parse of a
given input sentence. This methodology is usually based on the relative
frequencies of occurrences of the possible relations in a representative
corpus.
In the following text, we present an acquisition of training data for
the best analysis selection. The underlying mechanism is based on the
pruning constraints that automate the process of transformation of a de-
2when measured on 10 000 sentences from the DESAM corpus [PRS97]
70
3. SYNT – CZECH SYNTAX ANALYZER
Figure 3.1: Interaction of the NLP Centre parsing tools.
pendency treebank corpus. The results of the parsing are then compared
to running times of several referential parsing systems. The comparison
indicates that the synt system is fully able to compete with the best
current parsers.
This summary text on the synt parser is based on materials published
in [HKS02, HK05b, HK05a, HHK06, KKH06, HK06, KH07, HHKK07].
The core of the syntactic parser synt uses the meta-grammar formalism
(e.g. [Deb03, Kru03, HK05a]), which allows to specify complex
sentence constituent combination rules in a maintainable way. The composition
of the formalism resembles structures in Lexical functional grammars
[MK91] – the meta-grammar rules are expanded within the guidance
of combinatorial constructs plus contextual constraints and are supplemented
with additional actions and agreement tests. The described full
grammar development platform consists of the following components:
• the Czech morphological analyser ajka [Sed05]. This analyser provides
non-disambiguated output of plain text words. The ajka
system covers almost 400 000 Czech word lemmas, which generate
over 6 million word forms.
• the VerbaLex verb valency lexicon tools, see the Section 2.1. The
contextual constraints that safeguard the syntactic analysis of free
word order language need extra lexical-semantic information about
the sentence constituents. The creation of new verb valency lexicon
71
3. SYNT – CZECH SYNTAX ANALYZER
which contains so called complex valency frames has thus become
a part of the grammar development platform.
• the deep syntactic parser synt described in detail in the rest of this
chapter. The parser development concentrates on high coverage of
general corpus sentences (> 90 %). The parser is able to work with
several parsing algorithms (GLR parsing, top-down, bottom-up and
head-driven chart parsing). The synt parser also introduced a new
variant of the head-driven technique, the head-driven dependent
dot move algorithm (see the Section 3.2.2). This parsing technique
allows to parse natural language sentences within median time of
less than 0.1s/sentence.
• the user interface for linguists – the Grammar Development Workbench,
GDW, is presented in the Section 3.1.1. The parser synt
provides a command line interface, which is suitable for all forms of
batch processing. However, the grammar development conducted
by linguistic experts combines working with synt input parameters
(the meta-grammar, the corpus text, parameters guiding the analysis)
as well as thorough studying of synt outputs (tagged sentences,
syntactic or dependency trees, chart graphs). All these tasks can
be solved via GDW.
The grammar development platform includes several production tools
and resources as described above. The Figure 3.1 shows, how these tools
interact with each other. The Czech morphological analyser ajka and the
VerbaLex verb valency lexicon provide lexical information about words or
multi-word expressions in an input sentence. The syntactic parser synt
uses the data from lexicon and the meta-grammar to create syntactic
structures. The Grammar Development Workbench (GDW) is a user
interface able to control the parser. Main tasks of linguistic experts
working with GDW include running the parser, studying its outputs,
edit and test the meta-grammar and also build a tree bank of correct
derivation trees used for probability estimations and testing of grammar
changes.
72
3. SYNT – CZECH SYNTAX ANALYZER 3.1
3.1 The Grammar Development Process
The synt meta-grammar (denoted as grammar form G1) is carefully developed
by linguistic experts – currently it contains about 300 meta-rules
plus special generative constructs and selectional restrictions (see the
Section 3.2.1). For the actual parsing, this grammar form is automatically
expanded to one of the generated forms G2 or G3 (with about 3000
or 11000 rules), which describe the context free backbone supplemented
with additional tests and actions for capturing the context dependencies.
For supporting automatic consistency checks of the grammar development,
a new treebank3
of synt derivation trees is being developed. Any
changes in the meta-grammar then undergo an evaluation against this
treebank. All phases of this process are controlled with the Grammar
Development Workbench described below, which provides all necessary
functions to the linguistic experts working with the system.
The development of the parser goes hand in hand with the development
of the meta-grammar format. For decisions about the order of
implementation of proposed enhancement methods, different testing corpora
and treebanks are used for statistical estimates of the payoff of the
particular method. An example of such tree corpus utilization is presented
in the Section 3.5.
3.1.1 Grammar Development Workbench
All the above described tools and systems are presented to a user by
means of a graphical front-end tool – the Grammar Development Workbench
(GDW). Because most of the platform components (synt, ajka, ...)
are controlled only from command-line, the GDW graphical user interface
has been created to allow users comfortable and well arranged work
with the tools.
GDW has been developed within the diploma thesis of Radek Vykydal
[Vyk05] under the supervision of Aleš Horák. The system is still
under development and it is now extensively tested by linguists. GDW
consists of five modules:
• Gsynt – graphical user interface of the synt parser.
• TreeView – viewer for resulting syntactic trees.
3a collection of “trees”
73
3.1.1 3. SYNT – CZECH SYNTAX ANALYZER
Figure 3.2: The Gsynt module window showing an analysis of a sentence
from corpus.
• ChartView – browser of resulting chart structure.
• GrammarView – grammar forms viewer.
• TBAdmin – tree bank creation tool.
The linguistic work concentrates on the methodology of grammar development
with respect to real-world natural language texts. The most
important GDW tasks are enhancing the meta-grammar and creation of
a tree bank.
The process of building a treebank of correct syntactic trees consists
of the following steps: first, each input corpus sentence is analysed in the
Gsynt module. The Figure 3.2 shows the basic window with sentences
74
3. SYNT – CZECH SYNTAX ANALYZER 3.1.1
Figure 3.3: The resulting chart in the Grammar Development Workbench
tool.
from a corpus.4
The selected sentence is analyzed by the synt parser
and the output of the parser is displayed. Several additional information
about sentences is presented in the window.
The syntactic trees are displayed by the TreeView module. The resulting
syntactic trees can be filtered with several kinds of queries and
reduced by successive specification of pruning constraints. The one resulted
syntactic tree is added to the tree bank at the end.
Enhancing the meta-grammar lies in determining new rules for an
unaccepted sentence. By means of the ChartView module, see the Figure
3.3, a problematic sentence construct that is not covered by the metagrammar
can be located. Then the exact specification of the needed
4PDTB-1.0 [Haj98] in this case
75
3.2 3. SYNT – CZECH SYNTAX ANALYZER
rule(s) for the uncovered construct is determined and entered in the
GrammarView module, see the Figure 3.4.
We concentrate on the methodology of grammar development with
respect to real-world natural language texts. The particular procedures
include:
• building a treebank of correct syntactic trees, that is, for a correctly
analysed sentence:
– displaying the syntactic trees. Those trees are all consistent
with the imposed meta-grammar and they are ordered by the
computed tree ranks. However, the ordering is still premature
and the correct tree is often not in the first position yet.
– reduction of the number of syntactic trees with successive specification
of pruning constraints.
– adding the resulted syntactic tree to the treebank.
• enhancing the meta-grammar, that is determining new rules for an
unaccepted sentence:
– ChartView – helps locating the problematic sentence construct
that is not covered by the meta-grammar.
– GrammarView – for finding the exact specification of the missing
rule(s) for the uncovered construct.
For more detailed information about the GDW project see the project
documentation page [Vyk07].
3.2 New Meta-grammar Constructs in synt
The development of a grammar for syntactic analysis of natural language
texts is a tedious process which always tends to keep adding new rules for
capturing uncovered language phenomena up to the moment, where the
number of rules becomes hardly maintainable. Such “rule overload” is
solved with several competing approaches ranging from stochastic parsing
[Bod03], through various automatic learning algorithms [CD02] upto
loosening the strictness of the analysis with a shallow parsing algorithm
[vdBB02].
76
3. SYNT – CZECH SYNTAX ANALYZER 3.2.1
In the synt system, we have decided to apply the meta-grammar
concept with three consecutive grammar forms. The underlying analyzer
is an augmented head driven parser based on a context free backbone
with interleaved evaluation of contextual constraints.
3.2.1 The Meta-grammar Design
The meta-grammar concept in synt consists of three grammar forms denoted
as G1, G2 and G3. Human experts work with the meta-grammar
form (G1), which encompasses high-level generative constructs that reflect
the meta-level natural language phenomena like the word order constraints,
and enable to describe the language with a maintainable number
of rules. The meta-grammar serves as a base for the second grammar
form (G2) which comes into existence by expanding the constructs. This
grammar consists of context free rules equipped with feature agreement
tests and other contextual actions. The last phase of grammar induction
(G3) lies in the transformation of the tests into standard rules of the expanded
grammar with the actions remaining to guarantee the contextual
requirements.
The Figure 3.4 illustrates the use of the generative construct %list coord
case prep in G1, that produces two context free rules with pruning
actions in G2 and fourteen context free rules in G3. The grammars are
displayed by the GrammarView module, which is a part of the Grammar
Development Workbench environment described in the previous section.
The number of rules naturally grows in the direction G1 < G2 < G3.
The current numbers of rules in the three grammar forms are 236 in G1,
2741 in G2 and 11088 in G3, but the grammar is still being developed
and enhanced.
3.2.1.1 The Meta-grammar Form (G1)
The meta-grammar consists of global order constraints that safeguard the
succession of given terminals, special flags that impose particular restrictions
to given non-terminals and terminals on the right hand side (RHS)
and of constructs used to generate combinations of rule elements. Some
of these meta-grammar constructs have already been described in [HS00],
in this section we present a summary of all the constructs including the
latest additions.
77
3.2.1 3. SYNT – CZECH SYNTAX ANALYZER
Figure 3.4: Generative construct %list coord case prep in the grammar
G1 and the appropriate (generated) rules and actions in G2 and
G3.
The arrow mark in the rule is used for specification of the rule type
(->, -->, ==> or ===>). A little hint to the arrow form meaning can be
expressed by ‘the thicker and longer the arrow the more (complex) actions
are to be done in the rule translation’. The smallest arrow (->) denotes
an ordinary CFG transcription and the thick extra-long arrow (===>)
inserts possible inter-segments between the RHS constituents, checks the
correct order of enclitics and supplies several forms of the rule to make
the verb phrase into a full sentence.
The global constructs (%enclitic, %order and %merge actions) represent
universal simple regulators, which are used to inhibit some combinations
of terminals in rules, or which specify the actions that need some
special treatment in the meta-grammar form translation.
The main combining constructs in the meta-grammar G1 are order(),
78
3. SYNT – CZECH SYNTAX ANALYZER 3.2.1
rhs() and first(), which are used for generating variants of assortments
of given terminals and non-terminals.
/* budu se ptat - I will ask */
clause ===> order(VBU,R,VRI)
/* ktery ... - which ... */
relclause ===> first(relprongr) rhs(clause)
The order() construct generates all possible permutations of its components.
The first() and rhs() constructs are employed to implant
content of all the right hand sides of specified non-terminal to the rule.
The rhs(N) construct inserts all possible rewritings of the non-terminal
N. The resulting terms are then subject to standard constraints, enclitic
checking and inter-segment insertion. In some cases, one needs to force
a certain constituent to be the first non-terminal on the RHS. The construct
first(N) ensures that N is firmly tied to the beginning and can
neither be preceded by an inter-segment nor any other construct.
There are several generative constructs for defining rule templates to
simplify the creation and maintenance of the grammar. One group of such
constructs is formed by a set of %list * expressions, which automatically
produce new rules for a list of the given non-terminals either simply
concatenated or separated by comma and co-ordinative conjunctions.
A significant portion of the grammar is made up by verb group rules
(about 40 %). Therefore we have been seeking for an instrument that
would catch frequent repetitive constructions in verb groups. The obtained
addition is the %group keyword illustrated by the following exam-
ple:
%group verbP={
V: verb_rule_schema($@,"(#1)")
groupflag($1,"head"),
VR R: verb_rule_schema($@,"(#1 #2)")
groupflag($1,"head"),
}
/* ctu/ptam se - I am reading/I am asking */
clause ====> order(group(verbP), vi_list)
verb_rule_schema($@,"#2")
depends(getgroupflag($1,"head"), $2)
79
3.2.1 3. SYNT – CZECH SYNTAX ANALYZER
Here, the group verbP denotes two sets of non-terminals with the corresponding
actions that are substituted for the expression group(verbP)
on the RHS of the clause non-terminal. In order to be able to refer to
verb group members in the rules, where the group is used, any group
term can be assigned a flag (any string). By that flag an outside action
can refer to the term later with the getgroupflag construct.
As we can see in the example above, the verb rule schema action
(which defines the part of the verb group that forms a verbal object
in the successive logical analysis) appears in both parts of the RHS –
the group and the rule itself. For the logical analysis, the arguments
of this actions have to be gathered and merged into one resulting action.
This behaviour can be defined per-action with the global construct
%merge actions={verb rule schema}.
Many rules, e.g. those prescribing the structure of a clause, share
the same rule template — they have the same requirements for intersegments
filling and the enclitics order checking as well as the RHS term
combinations. To enable a global specification of such majority of rules,
we provide a rule template mechanism, which defines a pattern of each
such rule (the rule type and the RHS encapsulation with some generative
construct).
Some grammatical phenomena occur rarely in common texts. The
best way to capture this sparseness is to train rule probabilities on a
large data bank of derivation trees acquired from corpus sentences. Since
preparation of such corpus of adequate size (at least tens of thousands
of sentences) is a very expensive and tedious process, we have for now
overcome this difficulty with defining rule levels. Every rule without
level indication is of level 0. The higher the level, the less frequent the
appropriate grammatical phenomenon is, according to the guidance of
the grammarian. Rules of higher levels can be set on or off according to
the chosen level of the whole grammar.
3:np -> adj_group
propagate_case_number_gender($1)
In the above example the rule is of level 3, thus when we turn the grammar
level to at least 3, we allow adjective groups to form a separate
intersegment.5
When analyzing with grammar of level 0 the rule ‘np ->
5intersegment is a technical name of any constituent that may supplement the verb
phrase in a sentence.
80
3. SYNT – CZECH SYNTAX ANALYZER 3.2.1
adj group’ is not seen as a part of the grammar at all.
Apart from the common generative constructs, the meta-grammar
comprises feature tagging actions that specify certain local aspects of the
denoted (non-)terminal. One of these actions is the specification of the
head-dependent relations in the rule — the head() and depends() constructs
which allow to express the dependency links between rule terms.
3.2.1.2 The Second Grammar Form (G2)
As we have mentioned earlier, several pre-defined grammatical tests and
procedures are used in the description of context actions associated with
each grammatical rule of the system. The pruning actions include:
• grammatical case test for particular words and noun groups
• agreement test of case in prepositional construction
• agreement test of number and gender for relative pronouns
• agreement test of case, number and gender for noun groups
• type checking of logical constructions
np -> adj_group np
rule_schema($@, "lwtx(awtx(#1) and awtx(#2))")
rule_schema($@, "lwtx([[awt(#1),#2],x])")
The rule schema action presents a prescription for building a logical
construction out of the sub-constructions from the RHS. Each time a type
checking mechanism is applied and only the type-correct combinations are
passed through.
The G2 contextual actions propagate all and agree * and propagate
compute and propagate all relevant grammatical information from
the selected non-terminals on the RHS to the one on the left hand side
of the rule.
3.2.1.3 The Expanded Grammar Form (G3)
The feature agreement tests can be transformed into context free rules.
For instance in Czech, similar to other Slavic languages, we have 7 grammatical
cases (nominative, genitive, dative, accusative, vocative, locative
and instrumental), two numbers (singular and plural) and three genders
(masculine, feminine and neuter), in which masculine exists in two
81
3.2.2 3. SYNT – CZECH SYNTAX ANALYZER
Figure 3.5: Modules and data flow in the synt’s parsing librabry libkp
forms — animate and inanimate. Thus, e.g., we get 56 possible variants
for a full agreement between two constituents.
The G3 grammar represents such transformation of the G2 form with
the agreement actions interleaved into the rules. The actual parsing
process can work in two modes – either directly with the G2 grammar
and separate evaluation of agreement tests or with the G3 grammar,
where the agreement tests are inseparable part of the grammar rules. Our
experiments [HS00] show that the running times of these two approaches
are, in a general case, of nearly the same value – the G3 form performs
slightly better for more ambiguous sentences.
3.2.2 The Parsing Algorithm
We restrict our work to lexicalized grammars, where terminals can only
appear in lexical rules in the form of A → wi.6
This restriction allows us
to simplify the implementation and it also enables to separate a lexicon
from the grammar.7
The parsing module of synt, the libkp library developed
by Vladimír Kadlec in the MU NLP Centre, provides an efficient
implementation of standard parser tasks:
• syntactic analysis of sentences in natural language based on context
free grammars that can be large and highly ambiguous;
• efficient representation of derivation trees;
6i.e. preterminal category → terminal word.
7Actually, instead of using a static lexicon, we construct the lexical rules using the
Czech morphological analyser ajka [Sed05].
82
3. SYNT – CZECH SYNTAX ANALYZER 3.2.3
• pruning of the trees based on the application of contextual con-
straints;
• selecting n best trees based on computed edge rank values (e.g. the
frequency characteristics obtained from treebanks);
• visualization and printing of the parsing trees in a graphical form.
The parsing process consists of several steps. Firstly, the packed shared
forest [Ear70] is produced by the standard context free parsing algorithm.
Then the contextual constraints are applied. Finally, the best trees are
selected. The order of the last two steps can be reversed. We use this
multi-pass approach, because all these functions are implemented as plugins
that can be modified as needed or even substituted with other implementations.
For example, we have compared four different parsing algorithms
which use identical internal data structures (Earley’s top-down
and bottom-up chart parser [Ear70], head-driven chart parser [Kay89]
and Tomita’s GLR [Tom86]. All these implementations produce the same
structures, thus applying contextual constraints or selecting n best trees
can be shared among them. The data flow in the libkp library is shown
in the Figure 3.5.
The parsing technique of the synt system is based on the head driven
approach with improvements regarding the process of confirmation of
viable hypotheses. The HDddm (head driven with dependent dot move)
parsing technique [KS06] refers to the fact that the move of one “dot” in
the head driven parsing step is dependent on the opposite move of the
other one.
The efficiency of the parser depends to a considerable extent on the
choice of grammar rule heads. The current positions of heads in the
synt meta-grammar have been chosen experimentally and they are in
agreement with the conception of the leading constituent in traditional
Czech grammars.
3.2.3 Evaluation of Contextual Constraints
The contextual constraints (or actions) defined in the meta-grammar can
be divided into four groups:
1. rule-tied actions
2. agreement fulfillment constraints
3. post-processing actions
4. actions based on derivation tree
83
3.2.4 3. SYNT – CZECH SYNTAX ANALYZER
The rule-based edge ranking is solved on the first level by the rule-tied
actions, which also serve as rule parameterization modifiers.
Agreement fulfillment constraints serve as chart pruning actions and
they are used in generating the expanded grammar G3. The agreement
fulfillment constraints represent the functional constraints, whose processing
can be interleaved with that of phrasal constraints.
The post-processing actions are not triggered until the chart is already
completed. Actions on this level are used mainly for computation
of analysis transformations for a particular input sentence and particular
analysis. Some such computations (e.g. verb frame extraction, see
the Section 3.4) demand exponential resources for computation over the
whole chart structure. This problem is solved by splitting the calculation
process into the pruning part (run on the level of post-processing actions)
and the reordering part, that is postponed until the actions based
on derivation tree.
The actions that do not need to work with the whole chart structure
are run after the best or n best derivation trees are selected. These actions
are used, for example, for determination of possible verb frames within
the input sentence, which can produce a new ordering of the selected
trees, or for the logical analysis of the sentence.
The edge rank computations use information from several supporting
resources, such as a treebank of testing sentences or a list of verb valencies.
The latter one is currently the most promising resource since we are
working on the VerbaLex lexicon, see the Section 2.1, featuring syntactic
dependencies of sentence constituents, their semantic roles and links to
the corresponding Czech WordNet classes.
3.2.4 The synt Parser Implementation
In the synt, every grammar rule has zero, one or more semantic actions.
The actions are computed bottom-up serving the purpose of:
a) computing a value used by another action on the higher level;
b) throwing out incorrect derivation trees.
For example, the following grammar rule for genitive constructions in
Czech has three semantic actions:
npnl -> np np +0.0784671532846715
test_gen ( $$ $2 )
84
3. SYNT – CZECH SYNTAX ANALYZER 3.2.4
[0, 2, npnl] [0, 1, np] [1, 2, np]
value2
value1 value1
value2
value3
value4
List of values List of children
->
Figure 3.6: Example of the forest of values.
prop_all ( $$ $1 )
depends:1 ( $$ $1 $2 )
First line contains a grammar rule with its frequency obtained from a
treebank. The contextual constraints are listed on the lines bellow it.
The number 1 after the colon represents an internal classification of the
action. We can turn an evaluation of actions with specified type on and
off. The $$ parameter represents the return value. The $n parameter
is a variable where we store a value of n-th nonterminal of the rule.
Notice that the presented notation is not entered directly by users. It is
generated automatically from the meta-grammar G1.
3.2.4.1 The Representation of Values
[BBR87] showed that parsing is in general case NP-complete problem
if grammars are allowed to have agreement features. The pruning constraints
in synt are weaker than general feature structures. It allows
an efficient implementation with the following properties. A node in the
derivation tree has only limited number of values (e.g. the cardinality of
the set for noun groups in our system is max. 56 = 7 cases × 2 nouns ×
(3 + 1) genders [HS00]). In synt instead of usual pruning of the original
packed share forest, a new forest of values is built during the analysis.
The worst case time complexity for one node in the forest of values is
therefore 56δ
, where δ is the length of the longest right-hand side of
grammar rules. Note that this complexity is independent of the input
sentence.
85
3.2.4 3. SYNT – CZECH SYNTAX ANALYZER
Table 3.1: Results of running the synt parser on 10 000 corpus sentences
(from the DESAM corpus, running on Intel Xeon 2.2GHz).
#sentences 10000 sentences
#words 191034 words
Maximum sentence length 155 words
Minimum sentence length 2 words
Average sentences length 19.1 words
Time of CFG parsing 5 minutes 52 seconds
Time of evaluating constraints/actions 38 minutes 32 seconds
Overall time with freeing memory 46 minutes 9 seconds
Average #words per second 68.97
Size of the log file 1.2 GiB
#accepted sentences 9208
The values in the forest of values are linked with Earley’s items.8
An
item contains a single linked list of its values. Each value holds a single
linked list of its children. The child is one dimensional array of values.
This array represents one combination of values that leads to the parent
value. Notice that there can be more combinations of values that lead
to the same value. The i-th cell of the array contains a reference to a
value from i-th symbol on the RHS of the corresponding grammar rule.
In case that the i-th symbol is not used for computation of the parent
value, only the reference to Earley’s item from such cell is used.
The Figure 3.6 shows an example representing rule npnl -> np np
and containing three Earley’s items. Each item on the RHS of the rule
contains two values, value1 and value2. This gives us four possible combinations.
The semantic action computes from combinations value1 ×
value2 and value2 × value1 the same value value4. The combination
value2 × value2 was classified as incorrect (by the action), so it is not
displayed.
During this experiment, synt has achieved an average of 92.1 % coverage
on the DESAM corpus [PRS97] sentences with about 84 % cases
where the correct syntactic tree was present in the result. However, the
8For the description of Earley’s items see e.g. [AH02].
86
3. SYNT – CZECH SYNTAX ANALYZER 3.3
process of automatic determining of the one tree that is correct from the
point of view of a linguistic expert is still premature. An average time
of the analysis of one sentence from the corpus data was 0.28 seconds.9
More details of timing results can be found in the Table 3.1.
3.3 Best Analysis Selection – a Supervised
Construction of Pruning Constraints
A common approach to acquiring the statistical data for analysis of syntax
employs learning the values from a fully tagged tree bank training
corpus. Building such corpora is a tedious and expensive work and it requires
a team cooperation of linguists and computer scientists. At present
the only representative source of Czech tree bank data is the Prague Dependency
Treebank (PDT) [Haj04a], which includes dependency analyses
of more than 100 000 Czech sentences.
In order to be able to exploit the data from PDT in our non-dependency
analysis, we have supplemented the meta-grammar with dependency
specification for constituents. Thus the output of the analysis (one selected
derivation tree) can be presented in the form of pure dependency
tree.
This technique enables us to relate the output of our parser to the
PDT data. However, the profit of exploitation of the information from
the dependency structures can be higher than that and can run in an automatically
controlled environment. For this purpose, we use the mechanism
of pruning constraints. A set of strict limitations is given to the
syntactic analyser, which passes on just the compliant parses. The constraints
can be either supplied manually for particular sentence by linguists,
or obtained from the transformed dependency tree in PDT.
The transformation is driven by guidelines specified by linguists. These
guidelines contain the following information:
• afun — analytical function attribute from PDTB 1.0
• term — corresponding nonterminal or preterminal from the meta-
grammar
9The average running time includes generation of log file messages.
87
3.3 3. SYNT – CZECH SYNTAX ANALYZER
# afun term mtag lexit
Sb np k1
Sb Ap np
Obj np
Atr modif k2
AuxP pn
Table 3.2: Simplified example of transformation guidelines.
• mtag — morphological tag constraint
• lexit — lexical item constraint
The automatic procedure for generating the pruning constraints then
successively tries to match the analytical function attribute in the input
sentence with the records in the transformation guidelines. Each
match found is then checked for agreement in the particular morphological
tag and lexical item according to the given criteria (currently a
pattern matching based on text regular expressions). If all required fields
are consistent with the guidelines, the corresponding subtree is marked
as the specified nonterminal or preterminal from the meta-grammar. The
syntactic analysis with the pruning constraints applied then prunes those
parsing trees from the resulting chart that do not contain the requested
nonterminal or preterminal in that position.
If more than one records in the guidelines match, the first match is
applied. This mechanism allows to prefer the most specific records to
the general ones, which differ in the lexical item constraint or the morphological
tag constraint only, used e.g. in the differentiation of various
adverbial types:
# afun term mtag lexit
Adv np k1
Adv adv
The process of transformation guidelines preparation is divided into several
steps to assure the consistency of acquired pruning constraints. After
88
3. SYNT – CZECH SYNTAX ANALYZER 3.4
Table 3.3: Examples of the reduction of the number of parsing trees for
randomly selected sentences.
sentence # # words # analyses # pruned reduced
analyses to (%)
00214 30 3112504044 2146560 0.07
00217 3 2 2 100
00287 12 56 4 7
00308 7 10 6 60
00486 6 1 1 100
00599 35 44660 4872 11
00612 25 2369063760 1048896 0.04
00842 17 409920 6336 1.5
every change, the results are checked against a testing set of input sentences
and the differences are reported to the user for arbitration.
The integration of the pruning constraints obtained automatically
with the mechanism of transformation guidelines has shown to be very
efficient. The tedious work of the training data acquisition for the best
analysis selection algorithm has been substantially facilitated. Examples
of the reduction are displayed in the Table 3.3. The Table presents examples
of sentences which were randomly chosen from the set of 1000
sentences analysed first without pruning constraints and then with automatically
generated pruning constraints.
The average percentage of reduction on all the tested sentences has
achieved 30 % (see the Figure 3.7). The future work on the refinement
of transformation guidelines will be concentrated on further reduction of
the number of automatically pruned analysis trees.
3.4 Parsing with Verb Frame Information
In the case of syntactic analysis of a free word order language, as the
Czech language is, we need to exploit the language specific features for
obtaining the correct ordering of the resulting syntactic analyses. So far
the most advantageous approach is the one based upon valencies of the
89
3.4 3. SYNT – CZECH SYNTAX ANALYZER
0 10 20 30 40
020406080100
words in sentence
percentageofreduction
Figure 3.7: The dependency of the reduction (%) of the number of resulting
analyses on the number of words in the input sentence
verb phrase – a crucial concept in traditional linguistics, as well.
The ambiguity level in the syntactic analysis of free word order languages
suffers from the exponential explosion of the number of resulting
derivation trees. The main reasons for this combinatorial grow arise on
several levels of the sentence building process (prepositional attachment,
verb argument resolution, non-projectivity, ellipsis, morphological ambiguity,
etc.). Usual approaches [BN00] solve this situation with probabilistic
selection of the “best” tree or trees. The best trees are judged by
a probabilistic figure of merit. Our experiments show, that in the case of
really free word order languages (like Czech) the probabilistic measures,
without a combination with other techniques, are not able to cover the
complexity of the sentence syntax. That is why we need to exploit the
knowledge of the language specific features as described in [HS02].
The basic sentence frame is driven by the lexical characteristics of its
predicative construction based on the set of possible verb valencies of the
particular verb (see e.g. [TK98]). We have implemented the technique
of discovering the possible verb valencies from the resulting ambiguous
packed shared forest (stored in the parsing chart). This enables us to
work with verb valencies in two directions: a) using the VerbaLex valency
lexicon to prune impossible combination regarding the particular
90
3. SYNT – CZECH SYNTAX ANALYZER 3.4.1
verb, and b) automatically process large corpora for discovering possible
verb valencies that are missing in the lexicon. These valencies are then offered
to the linguistic expert for addition to VerbaLex. Similar approach
has been described in [GAL03], in which a partial parsing outputs were
used for obtaining the verb subcategorization information. Our approach
includes a full parsing of Czech sentence, which increases the credibility
of the verb frame information.
The part of the system dedicated to exploitation of information obtained
from a list of verb frames is necessary for solving the prepositional
attachment problem in particular. During the analysis of noun groups
and prepositional noun groups in the role of verb valencies in a given
input sentence one needs to be able to distinguish free adjuncts or modifiers
from obligatory valencies. The WordNet classes together with the
surface features in complex valency frames are directly used for setting
up a set of heuristic rules that determine whether a noun group found
in the sentence serves here as a free adjunct or not. The heuristics are
based on the lexico-semantic constraints derived from the VerbaLex links
to the EuroWordNet hypero-hyponymic hierarchy.
Certainly, when checking the valencies with VerbaLex, we discharge
the dependence on the surface order. Before the system confronts the
actual verb valencies from the input sentence with the list of valency
frames found in the lexicon, all the valency expressions are reordered.
By using the standard ordering of participants, the valency frames can
be handled as pure sets independent on the current position of verb arguments.
However, since VerbaLex contains an information about the
usual verb position within the frame, we promote the standard ordering
with increasing or decreasing the respective chart edge rank.
3.4.1 Automatic Extraction of Verb Frames from the
Packed Shared Forest
The verb frame extraction (VFE) process in the synt system is controlled
by the meta-grammar semantic actions. As we have described in the
Section 3.2.4.1, the parser builds a forest of values to represent a result
of the application of contextual constraints. The VFE actions are then
executed on a different level (see the Section 3.2.1) than the “usual”
actions, which allows us to apply VFE actions on the whole forest of
values.
91
3.4.2 3. SYNT – CZECH SYNTAX ANALYZER
First of all, we find all noun groups covered by the particular context
free rule. Then compatible groups10
are processed by the VFE action.
Notice, that this step suffers from a possible exponential time complexity
because we work with the derivation trees and not with the packed forest.
On the other hand our experiments show (see the Table 3.4) that in the
average case this is not a problem.
If the analyzed verb has a corresponding entry in VerbaLex, we try
to match the extracted frame with frames in the lexicon. An example
sentence analysis with the VFE process can be found in the Section 2.1.4.
We have measured the results of the first version of the automatic
verb frame extraction on 4117 sentences from the Czech corpus DESAM
[PRS97]. We have selected sentences which are analysed on the
rule level 0, i.e. sentences, which do not contain analytically difficult phenomena
like non-projectivity or adjective noun phrases. Even on those
sentences the number of possible valency frames can be quite high (see
the Table 3.4). However, if we work with intersections of those possible
valency frames, we can get a useful reduction of the number of resulting
derivation trees – see the examples described in the next Section.
3.4.2 Examples
The projection of the extracted valency frames to the corresponding VerbaLex
entry can be used as an effective pruning tool for decreasing the
number of successful derivation trees. As an example of such pruning,
we can have a look at the sentence
Pokud *uchazeči kurs rekvalifikace úspěšně absolvují*, budou
mít jistě uplatnění v zaměstnání.
(lit.) If *the candidates the retraining course successfully complete*,
they will certainly assert themselves in their job.
A graph of all closed ranges in the resulting chart of this sentence is
displayed in the Figure 3.8. The corresponding valency frame for the
verb ’absolvovat’ from VerbaLex:
absolvovat:1/graduate:1 (receive an academic degree upon completion
of one’s studies)
AG<person:1>obl
who nom VERB KNOW<education:1>obl
what acc
10Compatible in the term of derivation, i.e. groups within the same derivation tree.
92
3. SYNT – CZECH SYNTAX ANALYZER 3.4.2
Table 3.4: The results of verb frame extraction from the corpus DESAM.
Number of sentences:
count 4117
Number of words in sentence:
minimum 2.0
maximum 68.0
average 16.8
median 15.0
Number of discovered valency frames:
minimum 0
maximum 37080
average 380
median 11
Elapsed time:
minimum 0.00 s
maximum 274.98 s
average 6.86 s
median 0.07 s
Figure 3.8: A graph of all closed ranges in the resulting chart of the
sentence “Pokud uchazeči kurs rekvalifikace úspěšně absolvují, budou
mít jistě uplatnění v zaměstnání.” (see the Section 3.4.2).
93
3.5 3. SYNT – CZECH SYNTAX ANALYZER
There are 132 trees for that sentence in the parsing system synt. Due to
the free word order the sequence of sentence constituents is
subject (uchazeči/candidates) – object (kurs/course)
– verb (absolvují/complete).
According to the valency frame the subject is a noun in nominative and
the object is a noun in accusative. It is evident, that those elements
cannot form a nominal phrase. This constriction reduces the number of
trees to 24.
Another example is displayed in the following sentence:
Havel se radil s představiteli justice a vnitra o posílení práva.
Havel consulted with representatives of judiciary and home office
on the consolidation of the legal system.
The valency frame for the verb ’radit se’ from VerbaLex is:
radit se:1/debate:3 (discuss the pros and cons of an issue)
AG<person:1>obl
who nom VERB SOC<person:1>
opt
with+whom ins
ENT|ABS<entity:1|abstraction:1>
opt
about+what loc
The number of synt trees for this sentence is 2672. The part of sentence
with the preposition ’s’ (with) and a noun in instrumental and the part of
sentence with preposition ’o’ (on) and a noun in locative are necessarily
prepositional nominal phrases. The application of such limits in synt
allows a significant reduction of the number of trees to 18.
3.5 The Beautified Chart Method – Pruning
Technique Based on Linguistic Ad-
equacy
In this section, we present one of the methods used for automatic reduction
of the number of output syntactic trees – the method is named
the beautified chart. We describe the principles of the method as well as
its implementation in the synt system. We evaluate the impact of the
method on the numbers of the resulting derivation trees and particular
algorithms on the synt internal data structures.
94
3. SYNT – CZECH SYNTAX ANALYZER 3.5.1
3.5.1 Beautified Trees
Thanks to the nature of the synt meta-grammar concept (for keeping
maintainable number of meta-rules), the (generated) grammar of synt
contains a significant number of non-terminals that have low (or only
“technical”) impact to the linguistic interpretation of the results. These
non-terminals cause the fact that many differences between the resulting
derivation trees are also mostly technical. That is why we have introduced
a technique that rebuilds the derivation trees in a way, which keeps all the
linguistically relevant information and removes only the technical nodes
from the derivation trees to make so called beautified trees. The difference
between a regular tree and a beautified tree is shown in the Figure 3.9. As
we can see, the beautified trees are also simpler and more comfortable to
read than the regular output trees, which is however not the main reason
of the method implementation. The main reason lies in the reduction of
the number of output trees while keeping all the linguistically adequate
features of the trees.
As illustrated in the Figure 3.9, there are 4 basic modifications of the
derivation tree:
• rename – renames particular non-terminals in the resulting tree
(e.g. ss → sentence). This is a 1:1 operation.
• ignore – deletes particular (technical) non-terminals in the resulting
tree – the descendants of the deleted node are linked to its parent
node (e.g. intr, partnl)
• lists – rebuild the structured non-terminal lists to “flat” form (e.g.
the part list in the left sub-tree)
• importance – selection of one important non-terminal from each
non-terminal linear sequence and deleting the others (e.g. the np
non-terminal was selected from the sequence in the right hand part
of the tree in the Figure 3.9)
While transformation of a single tree is not a difficult algorithm, in the
synt parser, which is based on the head-driven chart parsing algorithm,
we need to “upgrade” this method to a transformation of the whole packed
shared forest of the results to keep a good running time and complexity
of the parsing.
95
3.5.1 3. SYNT – CZECH SYNTAX ANALYZER
A regular derivation tree:
The same tree as a “beautified” tree:
Figure 3.9: Regular tree vs. beautified tree
96
3. SYNT – CZECH SYNTAX ANALYZER 3.5.2
#sentences percentage
Accepted for testing 3393 67.86%
Excluded from testing 877 17.54%
Not accepted 730 14.60%
Total sentences 5 000 100.00%
Table 3.5: Results – all sentences.
The principal idea of the beautified chart method consists in an automatic
transformation of the synt internal data structures (chart, forest
of values) to new structures that will be able to generate (all) beautified
trees directly. The design of the used transformation operations follows
the requirement of reducing the structures, so the number of resulting
beautified trees is supposed to be (significantly) lower than numbers of
the original trees.
3.5.2 The Previous Estimate of the Effect of the Beautified
Chart Method
Before the actual implementation of the presented beautified chart
method, we have done a test aimed at finding out how much the (future)
method reduces the number of possible output trees. The task was to find
out how dramatically the number of derivation trees decreases, if there
is some correlation with the sentence length and other input variables.
In the test, we have randomly selected 5 000 Czech sentences of different
length (from 2 to 42 words). After processing the sentences with synt
the number of possible resulting beautified trees were computed, which
allows to estimate the asset of the method (for this computation we have
limited the number of output trees to 50 000). The new statistical quantity
Removed trees percentage was defined as 100 ∗ (nf − nb)/nf where
nf = number of full trees, nb = number of beautified trees.
According to the parsing results (see the Table 3.5), the sentences
were divided into three groups:
• Accepted for testing – the sentence was accepted by the parser,
number of regular trees was lower than 50 000 (Removed trees percentage
was computed only for these sentences).
97
3.5.2 3. SYNT – CZECH SYNTAX ANALYZER
Figure 3.10: Graphical representation of the Removed trees percentage
characteristics.
98
3. SYNT – CZECH SYNTAX ANALYZER 3.5.2
Results – short sentences (maximum 12 words):
#sentences percentage
Accepted for testing 1351 92.72%
Excluded from testing 0 0,00%
Not accepted 106 7.28%
Total sentences 1457 100.00%
Results – medium-length sentences (13 – 20 words):
#sentences percentage
Accepted for testing 1432 84.09%
Excluded from testing 86 5.05%
Not accepted 185 10.86%
Total sentences 1703 100.00%
Results – long sentences (more than 20 words):
#sentences percentage
Accepted for testing 610 33.15%
Excluded from testing 791 42.99%
Not accepted 439 23.86%
Total sentences 1840 100.00%
Table 3.6: Results – per sentence length (for short, medium and long
sentences).
• Excluded from testing – the sentence was accepted by the parser,
number of regular trees was higher than 50 000.
• Not accepted – the sentence was rejected by the parser.
The parser accepted 85 percent of the sample sentences, two thirds of
it were acceptable for our measurements. We can see in the Table 3.6
that the parser is (naturally) more successful on short sentences. For
sentences with more than 20 words, only 33 percent of sentences were
appropriate for our testing. It means that we had only little information
about the method’s effect on the long sentences. However, the actual
99
3.5.2 3. SYNT – CZECH SYNTAX ANALYZER
Number of sentences 3393
Average 21.513 %
Maximum 94.309 %
Minimum 0.000 %
Median 8.333 %
Variance 629.649 %
Standard deviation 25.093 %
Correlation between
number of words and 0.598
removed trees percentage
Correlation between
number of regular trees and 0.384
removed trees percentage
Table 3.7: Statistical characteristics of the Removed trees percentage.
Number of sentences 2042
Average 32.780 %
Maximum 94.309 %
Minimum 0.000 %
Median 33.333 %
Variance 629.574 %
Standard deviation 25.349 %
Table 3.8: Removed trees percentage - long sentences (more than 12
words).
100
3. SYNT – CZECH SYNTAX ANALYZER 3.5.3
Figure 3.11: Illustration of the Chart structure.
implementation showed that the effect of the method on long sentences
is even bigger.
Values of the Removed trees percentage oscillated between 0 and 94
percent, the average was around 22 percent. High variance and standard
deviation indicate that particular results could differ significantly. If only
long sentences are taken into consideration, the average and the median
were higher (see the Table 3.8). Correlation coefficients and diagrams in
the Figure 3.10 also indicated that there is a weak positive dependency
especially between the number of all trees and the removed trees percentage.
Then, we could expect higher efficiency at sentences with more
than 50 000 regular derivation trees. This were the numbers from the
pre-implementation estimation of the method’s asset.
3.5.3 The Beautified Chart Algorithm
In this section, we describe the internal synt structures, chart and forest
of values. We also describe the beautified chart transformations of these
structures.
Chart (denoted also as the “packed shared forest”) is the direct output
from the context free stage of the parsing. It is an oriented multigraph
where nodes are boundaries between words in the input sentence
and edges represent the part of the sentence that has been successfully
analysed by a certain grammar rule.
The example of the chart structure for an ambiguous Czech sentence
“Jak zemřel ?” (ambiguous: 1. How did he die?; 2. Did the yak (animal)
101
3.5.3 3. SYNT – CZECH SYNTAX ANALYZER
die?) is shown in the Figure 3.11.
In addition to this, there are so-called “parent” links between the
edges. These links are created in the process of context free analysis
and their task is to keep the information about the analysis run in the
chart structure. Following these links and the grammar rules incident to
particular edges, it is possible to generate all derivation trees from the
chart.
3.5.3.1 The Beautified Chart Algorithm on the Chart Struc-
ture
The beautified chart algorithm on the chart structure is based on removing
particular edges from the structure and redirecting the appropriate
“parent” links. The unwanted edges can be detected using the
non-terminals on the left side of the appropriate rules (only these nonterminals
are used for denominating nodes in the resulting derivation
trees). After deleting particular edge, its “parents” are linked to the
“child” of the deleted edge.
By implementing the beautified chart algorithm on the chart structure
we have encountered several complications. The most serious of them was
the problem that the chart structure (as the output of context free stage
of the parsing process) contains insufficient contextual information. Further,
we were unable to perform contextual actions computation on the
new “beautified chart” structure because this modified structure is not
compatible with the synt grammar. Thus, the beautified chart algorithm
on the chart structure is not compatible with contextual actions evaluation.
For this reason, we have decided to implement the beautified chart
method on the other synt internal structure, the forest of values.
3.5.3.2 The Beautified Chart Algorithm on the Forest of Values
Structure
After finishing the context free part of the parsing process, the contextual
actions computation needs to be performed on the resulting chart
structure. The application of contextual actions (e.g. agreement actions)
enables the parser to eliminate high number of incorrect derivation trees.
The main output from this process is the forest of values structure that
contains the “zipped” information of all successful derivation trees (i.e.
102
3. SYNT – CZECH SYNTAX ANALYZER 3.5.3
Figure 3.12: Illustration of the forest of values structure.
trees which fulfill all the contextual actions).
A simplified forest of values structure and the generated derivation
trees from this structure are illustrated in the Figure 3.12. The structure’s
basic unit is an object called value. Each value contains a link to an
edge from the chart structure. In the picture, the values are denoted
by the left non-terminal of the rule incident to this edge. Furthermore,
each value contains a pointer to a list of “children.” Each “child” is
an array of pointers to different value objects in the structure. We can
see that the length of this array matches to the number of descendants
of the appropriate node in the resulting tree, whereas the number of
“children” in the list matches to the number of successful sub-trees for
the appropriate value. We can also see that there can be more pointers on
a single value, which spares memory, but due to this fact, some beautified
chart modifications are more complicated to implement.
The beautified chart algorithm on the forest of values structure is
based on replacing particular pointers to the value objects by pointers to
their “children” elements. One step of the algorithm is illustrated in the
Figure 3.13. As we can see, the pointer to value v3 is replaced by two
pointers, v5, v6 for the first v3 “child” and v7, v8 for the second one.
The algorithm runs recursively from the bottom (it begins with the
values that are attached to the terminal rules). For each value object it
computes the list of “beautified children” by searching through its list of
children and making modifications like the one shown in the Figure 3.13.
103
3.5.3 3. SYNT – CZECH SYNTAX ANALYZER
Figure 3.13: The Beautified chart modification on the forest of values
structure
Figure 3.14: The Beautified chart algorithm
104
3. SYNT – CZECH SYNTAX ANALYZER 3.5.4
The basic idea of the algorithm is outlined in the Figure 3.14. The transformations
introduced in the Section 3.5.1 are implemented as follows:
• the rename transformation is straightforward, it consists in replacement
of particular non-terminal names.
• the ignore transformation consists in detection of the unwanted
values and their replacement as shown in the Figure 3.13. The
detection of the unwanted values can be made using the left nonterminals
of the grammar rules incident to particular value objects.
• the lists modification is similar to the previous one. The only thing
which is more difficult is that the value “parent” must be taken into
consideration by unwanted value objects detection.
• importance is the most complicated transformation. It is based
on comparing particular value objects (i.e. the appropriate nonterminals)
by their “importance” and deleting the less important
ones. In the algorithm, this comparison proceeds before adding the
“beautified children” to a particular value. In addition to that, we
have to check if both compared value objects are parts of a linear
sequence.
During the computation of the “beautified children” lists an additional
check must be performed to identify the (rather exceptional) situation
where a single (shared) value is deleted from just part of the output trees
set and it is kept in the other trees. In such case the value object has to
be copied and the lists of children reorganized.
3.5.4 The Beautified Chart Results
The time complexity of the described algorithm is linear with regard to
the size of the forest of values structure. According to the preliminary
results on a small set of Czech sentences, the real effect of the beautified
chart method is much bigger than we have estimated in the Section 3.5.2.
The average values of the Removed trees percentage quantity are higher
than 50 percent. The total number of trees removed by the beautified
chart method is even higher, around 90 percent of the total number of
regular trees.
105
3.6 3. SYNT – CZECH SYNTAX ANALYZER
Grammar CT Atis PT
Rules 24,456 4,592 15,039
Nonterminals 3,946 192 38
Terminals 1,032 357 47
Test sentences 162 98 30
Average Parses 5.4 940 more than 264
Grammar CT Atis PT
Table 3.9: Test grammars and test sentences.
We explain these high numbers by the fact that the effect of the
described method is much higher in case of sentences that were excluded
from testing in the Section 3.5.2. Our measurements show that the tree
number reduction is much higher for these sentences.
3.6 Parser Comparison Experiments
The effectivity comparison of different parsers and parsing techniques
brings a strong impulse to improving the actual implementations. We
have published two experiments [HKS02, HHKK07] comparing the synt
parser to Moore’s parser of English and a set of Czech parsers.
3.6.1 synt and Moore’s parser
In the first comparison, we have measured the running times of the synt
parser when analyzing the data provided at the Parser Comparison Data
web pages [MC07]. These web pages resulted from discussions at the
Efficiency in Large Scale Parsing Systems Workshop at COLING 2000,
where one of the main conclusions was the need for a bank of data for
standardization of parser benchmarking.
The best results reported on standard data sets (ATIS, CT and PT
grammars) before this experiment, were the comparison data by Robert
C. Moore [Moo00a]. In the package, only the testing grammars with
input sentences are at the disposal.
The basic characteristics of the testing grammars are presented in the
Table 3.9. A detailed description of these grammars is given in [Moo00b].
106
3. SYNT – CZECH SYNTAX ANALYZER 3.6.2
ATIS grammar, Moore’s LC3 + UTF 11.60
ATIS grammar, synt 4.19
CT grammar, Moore’s LC3 + UTF 3.10
CT grammar, synt 4.19
PT grammar, Moore’s LC3 + UTF 41.80
PT grammar, synt 17.75
Table 3.10: Running times comparison (in seconds)
The results of the parser comparison appear in the Table 3.10. The
values in the Table give the total CPU times in seconds required by the
parser to completely process the test set associated with the grammar.
Since we could not run the referential implementation of the respective
Moore’s parser on the same machine, the above mentioned times are not
fully comparable (we assume that our tests were run on a slightly faster
machine than that of Moore’s tests). The long running times for the data
of the CT grammar are caused by little ambiguity of the grammar, so
that the synt parsing technique, which is optimized for highly ambiguous
grammars, cannot display its strong suits.
3.6.2 Phrasal synt Compared with Dependency Pars-
ers
During last years a number of syntax parsers of the Czech language have
been implemented with the concentration to parsing of real texts (in
contrast to theoretical and demonstration parsers created in 80s and 90s
of the last century).
Some of those “real text parsers” came into existence in the team
around the Prague Dependency Treebank [Haj04a], we will call them the
Prague parsers although the best ones of them are variants of parsers of
British or American authors.11
The other set of compared parsers are variants of the parser designed
and implemented in the team of the NLP Centre at Masaryk university
in Brno (the synt parser), thus we call it the Brno parser in the context
of this comparison.
11see bellow for citations
107
3.6.2 3. SYNT – CZECH SYNTAX ANALYZER
Although all these parsers have been tested and used for several years
already, their implementations are running more or less independently
and no rigorous comparison of their effectiveness has been done yet.
Here we also try to formulate the main problems that have hindered
such comparison so far, then offer a solution of them and finally present
the results of the actual comparison. The Prague parsers have already
been compared and rated all together, so the novelty in this comparison
is the Brno parser synt that is based on completely different approaches
in contrast to the Prague parsers.
3.6.2.1 The Prague Parsers – Basic Characteristics
The set of dependency parsers selected and denoted as the Prague parsers
contains the following representatives:
McD McDonnald’s maximum spanning tree parser [McD06],
COL Collins’s parser adapted for PDT [HCRT99],
ZZ Žabokrtský’s rule-based dependency parser [HŽ06],
AN Holan’s parser ANALOG – it has no training phase and in the parsing
phase it searches in the training data for the most similar local
tree configuration [Hol05],
L2R, R2L, L2R3, R2L3 Holan’s pushdown parsers [Hol04],
CP Holan’s and Žabokrtský ’s combining parser [HŽ06],
The selection of Prague parsers was limited to the parsers contained
in CP, which is currently the parser with the best known results on
PDT including also other parsers like Hall and Novák’s corrective modeling
parser [HN05] or Nilsson, Nivre and Hall’s graph transformation
parser [NNH06]. These parsers were not included in the comparison,
since currently we do not have their results for all sentences of the testing
data set.
The Holan’s pushdown parsers, during their training phase, create a
set of premise-action rules, and apply it during the parsing phase. In the
training phase, the parser determines the sequence of actions which leads
to the correct tree for each sentence (in case of ambiguity, a pre-specified
preference ordering of the actions is used). During the parsing phase, in
108
3. SYNT – CZECH SYNTAX ANALYZER 3.6.2
each situation the parser chooses the premise-action pair with the highest
score. In the tests, we have measured four versions of the pushdown
parser: L2R – the basic pushdown parser (left to right), R2L – the parser
processing the sentences in reverse order, L23 and R23 – the parsers using
3-letter suffixes of the word forms instead of the morphological tags.
3.6.2.2 The Principal Differences of the Parsers
In contrast to the Prague parsers, the Brno parser synt is based only on
its meta-grammar, the parser does not have any training phase used to
learn the context dependencies of the input texts. All rules that guide
the analysis process are developed by linguistic and IT experts with all
the drawbacks it can bring (see the Section 3.6.2.2 for a description of
some of them).
The most principal difference between the two sets of parsers is, of
course, the underlying formalism and methodology of the parsing process.
This is however not the sort of difference that would cause problems
in the parser comparison. In this section, we will concentrate on the
problems arising with different input and output data structures, different
morphological and syntactic tagging and different presuppositions on the
input text that all need to be resolved before we can start with the real
comparison.
Q1: The Input Format The input of the Brno parser is either a
tagged text (from corpus or from other tagged source) with morphological
tags compatible with the tagset of the Czech morphological analyser
ajka [Sed05] or a plain text (divided into sentences), which is then processed
with ajka. Since ajka does not resolve ambiguities on the morphological
level,12
the Brno parser generally counts with the possibility
of ambiguous surface level tokens.
The Prague parsers use as their input also text split into individual
sentences, but with unambiguous morphological tags obtained from
Hajič’s morphological analyser and tagger [Haj04b].
Both morphological analysers (and thus both parser groups) use different
morphological tagging systems, which are not 1:1 translatable to
each other. However, the differences do not affect the most important
12ajka provides all possible combinations of morphological features of the input
words.
109
3.6.2 3. SYNT – CZECH SYNTAX ANALYZER
morphological features from the point of view of the syntactic analysis, so
we have used an automatic conversion with some information stripping.
Q2: Dependency Trees vs. Phrasal Trees The output of Prague
parsers is formed by dependency trees or graphs, whereas the output of
the Brno parser is basically formed by packed shared forest of phrasal
trees. The Brno parser includes the possibility of sorting the trees of
the shared forest and output n trees with the highest tree rank (a value
obtained as a combination of several “figures of merit,” see [HS02]).
This difference in the output format plus the fact that the Brno team
does not yet have a large testing tree bank of phrasal trees for measure-
ments13
was the cause of the biggest problems in the comparison. Since
the measurements had to be done on several thousands of sentences, we
have decided to use the PDT-2.0 tree bank14
[Haj04a]. Since this tree
bank provides only the dependency trees,15
we have decided to convert
them to phrasal trees using the Collin’s conversion tool [Col98] and then
measure the differences between the Brno parser output and this “phrasal
PDT-2.0” using the PARSEVAL and the Leaf-ancestor assessment techniques
(for more details see the Section 3.6.2.3).
Q3: One Resulting Tree vs. (Shared) Forest The output of the
Brno parser is formed by the resulting chart structure, which encompasses
a whole forest of derivation trees (all of them, however, have the same
root nonterminal that represents the successful analysis).
In order to be able to provide a comparison of this forest with the
one tree obtained from PDT 2.0 conversion procedure, we have for each
sentence extracted first 100 (or less) trees sorted according to the tree
rank. Each of these trees was then compared to the one from PDT and
the results are displayed with the following 3 numbers: a) best trees – one
tree from the set that is most similar to the desired tree is selected and
compared; b) first tree – the tree with the highest tree rank is selected
and compared; and c) average – the average of all trees is presented.
13Such tree bank of about 5 thousand phrasal trees is being prepared during this
year.
14The Prague Dependency Treebank, version 2.0, was created by the Institute of
Formal and Applied Linguistics, http://ufal.mff.cuni.cz.
15for more than 80 000 Czech sentences
110
3. SYNT – CZECH SYNTAX ANALYZER 3.6.2
Q4: Projective vs. Non-projective Trees The output of the Brno
parser is always in the form of projective trees, but a non-projective
phrase can, in some cases, be analysed with the mechanism of different
rule levels, that allow to handle special kinds of phrases. Nevertheless, the
Brno parser is not suitable for full analysis of non-projective sentences
at the moment. In the future, we will have to provide techniques like
corrections for non-projective parses described in [HN05].
On the other hand, the output of the Prague parsers, as a set of
dependency edges between words, can cross the word surface order without
problems. Thus it can represent projective as well as non-projective
sentences.
According to the Prague Dependency Treebank statistics, PDT contains
approximately 20 % of non-projective sentences. The sentences selected
for comparison are thus not limited to only projective sentences,
but the results are counted separately for projective and non-projective
sentences.
Q5: The Testing Data Set For the measuring and comparison of
parser effectiveness, we definitely need a set of data with syntactic annotation.
Such data are available for the dependency trees in PDT. The
tree bank has three parts – the training part (train), the testing part for
development (d-test) and the testing part for evaluation (e-test).
Since the Prague parsers use the first two sets for development and
because there is no such similar tree bank available for the phrasal trees
from the Brno parser, we have decided to use the PDT e-test part (approximately
10 thousand sentences) for the comparison and we try to
overcome the differences between the parser outputs.
One important difference regarding the testing data set is the fact
that the Brno parser does not have any training or learning phase – it is
purely grammar-rule based parser. The drawback of this fact is that the
Brno parser cannot automatically adapt to kinds of texts that were not
intended for analysis. The parser is designed to analyse only sentences of
the usual structure. Since the Czech language is a representative of free
word order languages, the parser allows an analysis of many possible word
combinations that can form even very “wild” Czech sentences, however, it
refuses to analyse texts like PDT sentences e-test#00017: “10 - 3 %” or etest#00554:
“Dítě 4 - 10 let : 1640 (Child 4–10 years:1640).” The Prague
111
3.6.2 3. SYNT – CZECH SYNTAX ANALYZER
Figure 3.15: The difference of the results with measuring on the converted
PDT trees and on the small Brno tree set
parsers, thanks to their stochastic nature, do not have any problems in
analysing such kinds of “sentences.”
3.6.2.3 The Results
We have used the PDT-2.0 e-test part, where the morphological tags were
automatically converted from the Prague tags to the ajka tags without
ambiguities. The e-test set contains approximately 10 000 syntactically
annotated dependency trees. To get trees comparable to Brno parser
output, we needed to convert these dependency trees to phrasal trees.
The conversion proceeded in two steps: first, the PDT-2.0 dependency
trees in PML format (the default format in PDT-2.0) were converted into
the CSTS format (earlier format of PDT) with PDT tool btred. Then,
the Collin’s conversion tool [Col98] was used to obtain PDT-2.0 phrasal
trees similar to the output of the Brno parser. The statistical features of
the e-test set are:
• 10148 sentences (173586 words)
• 7732 projective sentences
• 2416 non-projective sentences
• 87.7 % Brno parser coverage
Since the Brno parser does not provide output for all sentences in the etest
set (see the discussion in the Section 3.6.2.2), the actual comparison
112
3. SYNT – CZECH SYNTAX ANALYZER 3.6.2
Table 3.11: The results of the Prague parsers (precision = recall)
Parser all sentences non-projective projective
R2L 73.845% 69.823 % 75.735%
L2R 71.315% 67.297 % 73.204%
ANALOG 71.077% 66.625 % 73.169%
R2L3 61.648% 58.276 % 63.233%
L2R3 53.276% 49.672 % 64.912%
zz 75.931% 74.177 % 76.755%
col 80.905% 75.634 % 83.383%
MST 83.984% 82.230 % 84.809%
CP 85.85 % 83.434 % 86.979%
was run only on those sentences from e-test, that were accepted by the
Brno parser.
3.6.2.4 Measuring Techniques
The methodology for measuring the results of dependency parsing is usually
defined as computation of the precision and recall of the particular
dependency edges in the resulting graph/tree. These quantities are measured
for each lexical item and the result is then computed as an average
precision and average recall throughout the whole set.
In the case of phrasal trees we use the two following measures, PARSEVAL
and Leaf-ancestor assessment (LAA).
The PARSEVAL scheme utilizes only the bracketing information from
the parser output to compute three values:
• crossing bracket – the number of brackets in the tested analyzer’s
parse that cross the tree bank parse.
• recall – a ratio of the number of correct brackets in the analyzer’s
parse to the total number of brackets in the tree bank parse.
• precision– a ratio of the number of correct brackets in the analyzer’s
parse to the total number of brackets in the parse.
113
3.6.2 3. SYNT – CZECH SYNTAX ANALYZER
Table 3.12: The results of the Brno parser on the e-test set
cross-brackets precision recall LAA
all sentences
Best trees 4.473 60.228 % 60.645% 71.5 %
First trees 6.229 47.306 % 50.778% 69.1 %
Average 5.799 45.627 % 46.584% 69.0 %
projective sentences
Best trees 3.619 66.718 % 68.663% 73.1 %
First trees 5.289 53.028 % 57.630% 70.6 %
Average 4.942 50.859 % 52.552% 70.5 %
non-projective sentences
Best trees 7.251 39.615 % 35.727% 65.6 %
First trees 9.325 29.275 % 29.699% 63.5 %
Average 8.625 29.112 % 28.097% 63.3 %
Table 3.13: The results of the Brno and Prague parsers on the small Brno
tree set
cross-brackets precision recall LAA
Best trees 0.792 89.519 % 92.274% 97.2 %
First trees 2.132 70.849 % 74.358% 92.6 %
Average 2.311 63.330 % 64.453% 91.4 %
R2L 81.472%
L2R 81.634%
ANALOG 76.537%
R2L3 63.754%
L2R3 57.201%
zz 86.650%
col 90.129%
MST 89.889%
CP 91.912%
114
3. SYNT – CZECH SYNTAX ANALYZER 3.6.2
There are several known limitations [BSDH98] of the PARSEVAL technique.
It is not clear whether this metric can be used for comparing
parsers with different degrees of structural fineness since the score on
this metric is tightly related to the degree of the structural detail.
Also, in case of this comparison there is a problem with evaluating
PARSEVAL on non-projective sentences. Processing non-projective sentence,
the Collins’ conversion tool changes the word order in the resulting
phrasal tree, so the PARSEVAL numbers are expected to be bad.
The Leaf-ancestor assessment [Sam00, SB03] measure is more complicated
than PARSEVAL. It considers a lineage for each word in the
sentence, that is, the sequence of node-labels found on the path between
leaf and root nodes in the respective trees. The lineages are compared by
their edit distance, each of them having the score between 0 and 1. The
score of the whole sentence is then defined as the mean similarity of the
lineage-pairs for its respective leaves.
Since it considers not only boundaries between the phrases, the LAA
measure is supposed to be more objective than the PARSEVAL, even
at non-projective sentences. In this comparison we used the Geoffrey
Sampson’s LAA implementation [Sam07].
3.6.2.5 Problems and Discussion
Overall results of the Prague parsers testing are presented in the Table
3.11 in the form of percentage of correct dependencies for the whole
set of sentences, for non-projective and for projective only. The results
of the Brno parser on the whole testing set (with manual tagging from
PDT-2.0), e-test is displayed in the Table 3.12.
The experiment of comparing the results of parsers with dependency
and phrasal outputs has opened several problems that we have tried to
cope with. One of the main causes of these problems were the incompatibilities
between the “phrasal PDT” trees and phrasal trees from the Brno
parser. This was also the main source of low precision and recall of the
parser. In order to prove this thesis, we have (manually) prepared a small
set of phrasal trees16
in the form of the Brno parser trees and repeated
the measurements for this subset. The improvement of the results of the
Brno parser on this small subset may be seen in the Table 3.13 and in
the Figure 3.15.
16for 100 sentences randomly chosen from the e-test projective sentences
115
3.7 3. SYNT – CZECH SYNTAX ANALYZER
3.7 Further Development of synt
In this chapter, we have described several tools, methods and systems
that together form a platform for full syntactic analysis of free word
order natural language, the Czech language. The development of the
platform components is definitely not yet completed, however, already
at this stage the system offers very high coverage on common corpus
sentences while keeping the analysis time on the level suitable for batch
processing of large texts.
The methods of the best analysis selection algorithm show that the
parsing of inflectional languages calls for sensitive approaches to the evaluation
of the appropriate figures of merit. The acquisition of these output
arranging quantities is based on a representative training data set. The
method of pruning constraints described in this text enables to automate
the process of treebank corpus transformation.
The integration of the presented methods to the parsing system has
no destructive impact on the efficiency of the parser. This is documented
by the comparison of the running times. In the described experiment
with Moore’s parsers, our system outperforms the results of the best
referential parsing system on highly ambiguous grammars, for which it
is optimized. In the second experiment, we have described a thorough
comparison of the techniques and outputs of the two groups of parsers of
the Czech language – the stochastic dependency Prague parsers and the
meta-grammar phrasal Brno parser. We have summarized and discussed
the problems of a comparison of such different approaches and we have
presented the measured results of the experiment. The results show that
the Prague stochastic parser are better for general textual data, which
do not have to follow (Czech) grammatical structures. However, it is not
easy to give such conclusion for proper sentences.
In the future development, we would like to repeat these tests on
another set of input data, namely on the prepared Brno phrasal tree
bank. The question is whether this different testing set will shuffle the
table of results significantly or it will stay more or less the same.
The presented parsing system synt has already proven its abilities in
analysis of running texts in both speed and coverage of various sentence
types. With continuous development of the grammar in order to obtain
coverage of all the necessary language phenomena, we obtain a quality
general purpose system for deep syntactic analysis of natural language
116
3. SYNT – CZECH SYNTAX ANALYZER 3.7
texts even for language with such extent of non-analytical features as the
Czech language is.
The current development of the system lies mainly in ordering of the
obtained analyzes with the usage of language specific features such as the
complex valency frames in the VerbaLex valency lexicon. The preliminary
results of the exploitation of VerbaLex in the syntactic analysis of Czech
are very promising and the precision of the analysis grows significantly.
We believe that with enlarging the lexicon to a representative number of
Czech verbs the synt system will be able to detect the correct derivation
tree in many cases which were unsolvable so far.
In the future development, we want to explore and implement methods
aimed at reducing the number of output derivation trees and the
best analysis selection. Systematic dealing with the syntax ambiguity of
the Czech language is the main task for future development of the synt
parsing system.
117
3.7 3. SYNT – CZECH SYNTAX ANALYZER
118
Chapter 4
Transparent Intensional
Logic as a Way to
Semantics
Knowledge representation and reasoning systems usually aim at applications
where the speed of responses to questions force using an underlying
formalism with limited expressive power. With this approach it is feasible
to describe a selected field of interest with a possibly huge amount of
facts and reuse this “knowledge” in an expert system tool.
Another approach, with a very small number of real implementations,
is based on the fact that most of the human knowledge is encoded in the
form of natural (non-artificial) language. Thus a straightforward way to
handle such information is to build a system capable of analyzing the
sentences directly with a machine parseable output and a connection to
an automatic inference machine.
A system that will fulfil such requirements is based on Tichý’s Transparent
Intensional Logic (TIL, [Tic88]). The work [Hor02] describes the
logical system based on the Normal Translation Algorithm (NTA) for
TIL. It follows the approach of full natural language analysis in the form
of normalized TIL constructions (denoted as concepts) as a meaning
bearer for natural language expressions and assertions.
In this chapter, the types of lexicons necessary for the TIL logical
119
4. TRANSPARENT INTENSIONAL LOGIC
analysis will be described. We will show the algorithm for analysing
the TIL verbal object as the core of a clause construction within the
sentence analysis. Examples of verb frame analysis for Czech words will
be presented. We also depict a way of enriching the lexicon entries as
combinations of the descriptions of lexical units as they are developed
within the area of lexical semantics (e.g. WordNet) with logical analysis of
sentence meanings worked out within the Transparent Intensional Logic
framework.
In the following text, we propose a method to integrate the logical
analysis of sentences with the linguistic approach to semantics, exploiting
the complex valency frames (CVFs) in the VerbaLex verb valency lexicon
presented in the Section 2.1. Since so far these two ways of description,
namely the logical and linguistic one, have been treated separately, the
task we set is to propose a method of their interrelation and coordination.
Needless to say that both ways of description of verb semantics are useful.
Hence we are going to show how to combine a logical description using
mostly terms like types, individuals, classes, relations, propositions, or,
in general, constructions of these entities, with the linguistic framework
capturing the idiosyncratic semantic features of verbs. In the Section 4.3
we adduce examples of the analysis of selected English and Czech verbs
for which the above mentioned integration has been proposed.
A question may be asked why we do not exploit first order predicate
logic (PL1) where some of the presented problems have already been
explored with PL1 used to represent logical forms. Nevertheless, it is a
well established fact that PL1 is not able to handle systematically the
phenomena like intensionality, belief attitudes, grammatical tenses and
modalities (modal verbs and modal particles in natural language). On
the other hand, since TIL works with techniques designed for capturing
the natural language meaning these problems either do not arise or they
can be solved in an intuitive way in TIL (see [Tic88]).
The following text further extends the work described in [Hor02, chapters
5 and 6] and summarizes the results published in [Hor04, HP04,
HPDM07, GH07a].
120
4. TRANSPARENT INTENSIONAL LOGIC 4.1
4.1 Overview of the Transparent Intensional
Logic
The details of the design of TIL have been described in several books,
see e.g. [Tic88, Tic04, Mat04, DJM08, Hor02]. Here we will present just
a very short overview.
TIL has come out of the ideas of Gottlob Frege, mostly following his
fundamental predicament about meaning, known as the Compositionality
Principle: the meaning of a compound is a function of the meanings of
its constituents.
One of the main assets, TIL brings to natural language analysis, is its
(hyper)intensionalistic feature, which is enabled through the possibility
of using and mentioning objects of complex types, i.e. not only objects
of the universe (individuals), but also functions of individuals, functions
of functions of individuals, etc.
The mechanism of intensions and extensions can be displayed with
the following example of two expressions: the dean of the Faculty of
Informatics and Jiří Zlatuška. The latter expression is a label of an
individual, which is independent on the state of the world and on the
time moment. In the case of the first expression the situation is different.
The dependency of this expression on time moment is clear: the dean of
the Faculty of Informatics was not the same in every time moment of the
history. It was, for example, also Luděk Matyska or nobody before 1994
when the faculty was established. The fact that Jiří Zlatuška is now the
dean of the Faculty of Informatics is only one of many possibilities, that
could have happened. We thus analyse this fact with an intension – a
function returning a truth value according to a selected possible world and
time moment. In TIL, objects depending on the state of the world and
on time moment are called intensions, other objects are extensions. Nonintensionalistic
logic systems, like PL1 which is often used to represent
real world facts, have no means to work with intensions.
Another powerful feature of TIL is the extended type hierarchy – TIL
works with objects of higher order type. That means we can represent
belief sentences in the same way as any other object and that the representations
do not lead to any inconsistencies in further processing. A
good exemplification is a statement about numerical calculation: if we
say Paul counts the sinus of π/2, a first order type representation would
121
4.1.1 4. TRANSPARENT INTENSIONAL LOGIC
make the expression the sinus of π/2 indistinguishable from the number
1. However, this obviously does not reflect the real meaning of the sentence
– Paul does not really need to know what is the correct result of
the sinus of π/2 and thus he may not even think about any particular
number here. The TIL analysis of this sentences solves this by relating
Paul to the construction of the respective mathematical object.
4.1.1 TIL Types
The type system of TIL was based on the theory of types introduced by
Church and was then further extended by Tichý [Tic82]. In TIL, every
object (as a representative of an entity described by the analyzed natural
language expression) has its type which is defined over a firmly set type
base.
Every TIL object is assigned either one of the basic types, or a type
that is formed by a mapping from one type to another type, or a higherorder
type. Within this framework, we can obtain an infinitely nested
hierarchy of types, i.e. mappings of basic types, mappings of mappings,
mappings of mappings of mappings, etc. Nevertheless, how difficult soever
the mapping is, the object of the respective type is still “flat.” The
flatness of mappings is predicated upon the way mappings are treated –
as collections of (n+1)-tuples (in case of an n-adic mapping). This means
that the mapping is (virtually) represented by a table of values without
any possibility to find out the way (a procedure) which leads to those
values. This is also one of the reasons why mere mappings cannot serve
as surrogates for meaning – mappings lack a constructing structure.
For capturing this structural property, we do not work directly with
TIL objects. We rather reference them by the schema describing, how
(in correspondence with the natural language expression) the TIL object
is constructed.
A construction is usually displayed in the typed λ-calculus notation,
as a λ-term, but the real construction is still the procedure described
with the term, not the term itself.
Variables are the only simple constructions. There are three modes of
forming constructions (from non-constructions and other constructions):
trivialization, composition and closure. Classes of all constructions form
another basic types in the hierarchy.
The idea of logical analysis of natural language with TIL lies in the
presupposition that every language has a definite intensional base – a
122
4. TRANSPARENT INTENSIONAL LOGIC 4.1.1
collection of fundamental properties of objects (not only individual properties
of type (oι)τω but any world and time depending relations among
objects like colors, heights, attitudes, . . . ), that are capable (without any
need for other extra techniques) of describing a (thinkable) state of the
world.
In TIL, such an intensional base of a natural language is rigorously
explicated in an epistemic framework, i.e. a typed system based on a
set of four basic types of order 1 (simple types) and collections of all
constructions of order n as types of order n + 1 (higher-order type).
These higher-order objects allow us to work not only with extensions
and intensions as representatives of real-world objects, but also with so
called hyperintensions,1
i.e. the constructions of intensions in the role of
TIL objects.
The four simple types form a type base and are denoted with letters
o, ι, τ and ω, with the following interpretations:
o is a set of two items representing the truth-values True (T) and False
(F). These two objects behave exactly the same as their counterparts
in the standard predicate logic especially in combination with
standard logic operations such as conjunction, disjunction, implication
or negation. These predicate logic operations can be represented
as objects of type (oo) or (ooo), i.e. functions with one or
two arguments of type o returning a value of type o.
ι is a class of individuals. The designation “individual” must not entice
us to imagine the members of this class as beings with all their
properties. In TIL, the notion of an individual is best interpreted
as a mean of a numerical identification of a (type unstructured2
)
entity. Any individual properties are ascribed to an object of type ι
only by means of asserting a statement that contains the ascribing
as its part – in the proposition ‘The Earth is flat,’ we use the ι-object
Earth as an identification of an entity that is ascribed the
property being flat (Flat)
λwλt[FlatwtEarth]
1The term “hyperintensional” has been introduced by Max Cresswell in [Cre75].
See also [Cre85].
2i.e. whose type is not decomposable as a function to the types of the functional
arguments and the type of its return value.
123
4.2 4. TRANSPARENT INTENSIONAL LOGIC
The individual Earth itself is carrying no a priori properties, it
serves as an identifier of a further unspecified object and is mainly
used for references to this object.
τ is a class of time moments. Due to the continuity feature of this class,
it may be regarded identical with the class of real numbers in case
we specify a fixed zero point and a unit. Functions working with
arguments of type τ are usually used for expressing the temporal
dependency of an entity.
ω is a class of possible worlds. Its members are intended for a transparent
representation of modal dependency of described objects.
A possible world, in accordance with the Leibnizian claims, is in
TIL defined as a determination system that, for each of the intuitively,
pre-theoretically given features from the intensional base,
contains an assignment of all possible (consistent) distributions of
those features.
Thus the selection (anchoring) of one such possible world allows
us to work with values that are dependent on the actual world
without the need of knowing exactly what the actual world is. This
is necessary, since the knowledge of the actual world itself would
bring severe inconsistencies into any theory that would rely upon
it.
The non-basic types are classes of special entities build upon this base.
They form an infinite hierarchy. The classes are formed by mappings
from and into the initial objects, mappings from and into those mappings,
etc. For every type a countable set is given, whose members are called
variables.
The types can be also viewed as classes of TIL objects, i.e., if ξ is a
type, A ∈ ξ then A is called a ξ-object. In TIL, we often work with objects
of particular types. The most frequent derived types are described in the
Table 4.1.
4.2 The Logical Analysis of a Sentence
For the purpose of obtaining the logical analysis of a sentence within the
Normal Translation Algorithm (NTA, [Hor02]), we describe the proce-
124
4. TRANSPARENT INTENSIONAL LOGIC 4.2
Table 4.1: The most frequent derived types in TIL (ξ in the table stands
for any type). Also the short notations are displayed.
Type Object
((oτ)ω),
oτω, π
a proposition – an assertion whose truth-value depends on
the world and time. Example: ‘John is old.’
((ιτ)ω), ιτω an intensional role – an individual office that may be engaged
by different individuals in different worlds or times.
Example: ‘the rector of Masaryk University’
(((oι)τ)ω),
(oι)τω
a property – an atomic intensional feature that an individual
either has or has not according to a chosen world
and time. Such feature is then represented by a class (its
characteristic function) of those individuals that have the
feature in a certain world and time moment. Example:
‘being rounded’
(ξτ), ξτ a ξ-chronology – a mapping that specifies the flow of
changes of a ξ-object in time. Example: ‘yesterday’ is
a time interval of 24 hours that ends at the last midnight,
thus it represents a different interval every day.
((ξτ)ω), ξτω an intension – expresses the dependency of the related ξobject
on the selected possible world and time moment.
The application of an intension to the world and time (i.e.
composition of a ξτω-object with argument of type ω and
then with argument of type τ resulting in a construction of
a ξ-object) is called the intensional descent. If ξ is not itself
an intension, it is called an extension and represents an entity
whose value does not change with world and time. All
mathematical objects (numbers, operations, axioms) correspond
to extensions.
125
4.2.1 4. TRANSPARENT INTENSIONAL LOGIC
dure of translation of a textual form of a natural language sentence (in
Czech) into the corresponding construction of Transparent Intensional
Logic, which then describes the logical meaning of the sentence.
In some cases, the resulting sentence construction is not pragmatically
anchored, i.e. it contains free variables, which reference the pragmatic
situation of the discourse as a whole. Before using such sentence in inference,
the pragmatic meaning of the sentence must be acquired with filling
the values of the free variables that represent personal pronouns, proper
names or discourse links. However, the discourse analysis is beyond the
scope of NTA and is not further discussed in this text.
In the following sections, we concentrate on the clause construction
with the usage of three lexicons:
• the lexicon of tokens
• the lexicon of verb frames
• the lexicon of functional items such as prepositions or conjunctions.
The content and format of these lexicons is exemplified further in the
text.
4.2.1 Verb Frame Analysis
The TIL type of the object, that is denoted by a verb in the finite form,
can be derived from the actual verb frame instantiated in the sentence.
Each of the verb arguments may be assigned a different type from the
lexicon of verb frames. In the VerbaLex lexicon of verb valencies for
Czech (see the Section 2.1), we record the syntactic surface structure of
the sentence constituents as well as their semantic role. During the logical
analysis in TIL, we need to identify yet another level of the denotation
of a verb argument – its meaning function. On this level, we enter the
construction of the TIL object represented by the corresponding natural
language expression.
The distinction of the three levels of verb frame representation may be
demonstrated on the example of the verb ‘brát’3
with a valency ‘někomu
něco.’4
The three levels then can look like5
3‘to take’
4‘something from somebody’
5At this level of analysis, the Agent (subject) is not processed. Its processing is
postponed to the clause level analysis.
126
4. TRANSPARENT INTENSIONAL LOGIC 4.2.1
1. syntactic surface structure:
brát
někomuhuman.NP, dat., no prep.
něconon-human.NP, accus., no prep.
This level reflects the properties of constituents that can be derived
following the morphological and syntactic analysis of the sentence.
2. semantic function:
AG<person:1> brát PAT<person:1> POS<possession:2>
The semantic function denotes the role of the verb arguments in
the activity expressed by the verb – PAT, the one that is referenced
by the verb as the receiver of the activity, and POS, the PAT’s
possession that is taken away.
3. meaning function:
brát/(o(oπ)(oπ))ω ιι
x . . . ι
y . . . ι: swty, s . . . (oι)τω
On this level, we try to find the construction of the object that
is represented by the corresponding constituent – x . . . ι, a specific
individual, and the other y . . . ι : swty, s . . . (oι)τω, an individual
from a class of individuals or an individual with a specified property.
In the analysis of the verb valency frame in the NTA, we need to find the
appropriate translation from the syntactic structure to the meaning function.
The particular construction and type that appears in the resulting
sentence analysis depends on (at least):
1. the actual input lexical items the constituent consists of – their
analysis has to be found in the lexicon of tokens.
2. the context – the lexicon often offers more than one possible analysis
of the lexical item. However, on the upper level the surrounding
lexical items may provide more details to the specification of the
subject and so allow to select only the appropriate analyses of the
item.
127
4.2.1 4. TRANSPARENT INTENSIONAL LOGIC
Table 4.2: Examples of a noun analysis.
Noun Analysis Description
pes, člověk *
x . . . ι: peswtx,
pes/(oι)τω
an individual from the class of
individuals – such x for which
peswtx holds
prezident *
prezident/ιτω an individual role
volitelnost *
volitelnost/
(oιτω)τω
a property of an individual role
výška,
hmotnost *
výška/(τι)τω a quantity
výrok,
tvrzení *
p . . . ∗π:
výrokwtp,
výrok/(o∗π)τω
a construction of a proposition
from the class of constructions
of a proposition
válka, smích,
zvonění *
válka/(o(oπ))ω a class of episodes – an activity
(with some time persistence)
that directly corresponds to an
episodic verb
leden,
podzim *
leden/(o(oτ)) classes of time moments – time
intervals specified by month or
season.
*‘dog’, ‘human’; ‘president’; ‘eligibility’; ‘height’, ‘weight’; ‘statement’,
‘assertion’; ‘war’, ‘laughter’, ‘ringing’; ‘January’, ‘autumn’
The basic guide-post for the list of valencies of Czech verbs that keeps
the syntactic structure of the verb valency should route the translation
of a valency expression (i.e. a specification of a verb argument) in the
following way:
a noun group (with/without preposition).
A noun phrase is usually formed by a core, a noun, which is preceded
by adjuncts in the form of an adjective, a pronoun or a numeral
or a combination of such items. In the simplest case, the noun
phrase consists of just one noun, whose analysis is found in the
128
4. TRANSPARENT INTENSIONAL LOGIC 4.2.1
lexicon. Examples of common analyses of a noun are presented in
the Table 4.2.
an adverbial phrase
The constructions of adverbial phrases usually works as a modifier
of the verbal object of the verb and is not described here in detail.
a subordinate clause
The sentence building includes the description of analysis of relative
and other subordinate clauses by means of clause combinations with
the conjunction object as its functor.
an infinitive
The infinitive form of a verb in the position of a verb argument
is analysed as the world-instantiated verbal object (object of type
o(oπ)(oπ)) of the corresponding verb.
Following these guidelines, the current lexicon of verb frames provides
the system with the information about:
• verb lemma
• surface verb frame – information about morphological and syntactic
features of verb arguments
• the TIL types of the arguments of the verbal object
• the schema of the verbal object construction
An example of the current version of a lexicon entry for the verb ‘brát’
with valency frame ‘někomu něco’ (to take something from somebody) is
presented in the Figure 4.1.
For an automatic verb object type assignment based on the VerbaLex
lexicon see the detailed description in the Section 4.3, where we show
how CVFs describe the surface valencies of verbs (i.e. their respective
morphological cases in Czech) as well as the semantics of their predicateargument
structure. Concerning the latter we make use of the deep semantic
roles expressed by two-level labels based partly on the Top Ontology
(EuroWordNet [VB+
98]) and partly on the selected literals from
Princeton WordNet [Mil90].
129
4.2.2 4. TRANSPARENT INTENSIONAL LOGIC
; lemma
brát
; encoded surface valency frame with TIL types of arguments (i=ι)
hPc3t{i}-hTc4t{i}
; the verbal object schema
:exists: V(v):V(v) :and: V(v)=[[#0,try(#1),try(#2)],V(w)]
Figure 4.1: Lexicon entry for the verb frame ‘brát někomu něco’ (to take
something from somebody).
4.2.2 The Sentence Analysis
In this stage, we have information about the logical analysis of a verb
group with its arguments and adverbial modifiers. What remains to be
specified, is the basic guide-post that suggests the best order in which all
the partial analyses of phrases, clauses and sentence should proceed.
We suppose that the input state of the logical analysis is formed
by an already disambiguated (uniquely identified) syntactic derivation
tree. Hence, the logical analysis may run either after the end of the
syntactic analysis of the input sentence, or as well in parallel with it,
in which case the necessary procedures perform as certain contextual actions,
which work over the possible combinations of the (locally) analysed
constituents. The asset of such parallel approach lies in its capability to
prune analyses which are type-inconsistent, e.g. if the verb expects an individual
as its argument, the type checking mechanism would not allow
a proposition to take this place. However, the cases where such pruning
may reduce the extent of the syntactic analysis are quite rare (remember,
this pruning applies only on sentences which are correct in their syntax
but inconsistent in the types of their constituents) or they can be substituted
with the verb frame analysis only. The drawback of the parallel
analysis lies also in the time and space spent on the overabundant logical
analysis of those subtrees that are not part of the resulting derivation
tree (i.e., in the parallel analysis, we cannot cast away any subtree that
is successful “so far,” even if it may be ruled out within the successive
analysis).
That is why, we have chosen the logical analysis to run within the
fourth group of our meta-grammar actions – the actions based on deriva-
130
4. TRANSPARENT INTENSIONAL LOGIC 4.2.2
tion tree.6
The process proceeds in certain (time) successive steps, which
are summarized in the following paragraphs.
4.2.2.1 Lexicon of Tokens
The logical analysis starts to build the construction of the whole sentence
from inside, i.e., in concordance with the Compositionality Principle,
the meaning of the compound is constructed as the meaning of its
constituents. Therefore, the first step must necessarily run in the lowest
part of the derivation tree – the analysis of the input lexical items. In this
step, we have not much choice other than to look up the proper analysis
(analyses) of the lexical items in the lexicon of tokens. An example of
entries in the lexicon of tokens is:
/* type definition: (oι)τω = property (of an individual) */
#T PROP (((oi)t)w)
/* lemma */
pečený
/* schema – object trivialization, 0
pečený . . . (oι)τω */
O(pečený/T PROP)
/* lemma */
kuře
/* schema – property ascription, (∃i . . . ι)[0
kuřewt i] */
:exists: V(i/i):V(i) :and: [awt(O(kuře/T PROP)),V(i)]
Here, the word ‘pečený’ (roasted) is analysed as a property of individuals
and the word ‘kuře’ (chicken) is given the analysis of an individual bearing
a specific property (to be a chicken).
The lexicon can supply some wild-card values based on the grammatical
category of the lexical item, but in such case we risk the possibility
of incorrect type assignment (e.g. the word ‘výška’ (height) cannot be
analysed as an individual). Hence, as a result of this part, we receive
the type of each lexical item as well as a schema of its working with
other (dependent) constituents (e.g. a conjunction is accompanied with
the schema of the relevant clause (propositions) as its arguments). Such
a lexical item that expects some arguments to be meaningful is called a
functional lexical item.
6see the Section 3.2.3.
131
4.2.2 4. TRANSPARENT INTENSIONAL LOGIC
prep_noun_phrase -> prep noun_phrase
agree_case_and_propagate($1, $2)
depends($1,$2)
add_prep_ngroup($2)
rule_schema($@,"lwt([awt(#1),try(#2)])")
noun_phrase -> left_modif noun_phrase
agree_case_number_gender_and_propagate($1, $2)
depends($2,$1)
rule_schema($@,"lwtx(awtx(#1) and awtx(#2))")
rule_schema($@,"lwtx([[awt(#1),#2],x])")
Figure 4.2: Examples of meta-grammar rules for prepositional noun
phrase and noun phrase with multiple rule schemata. The actual choice
of the proper rule schema is guided by the type checking system during
the running time of the analysis.
4.2.2.2 Rule Schemata
The analysis then moves up the derivation tree, rule by rule. Each rule
is supplemented with a similar schema as the functional lexical items, a
schema that tells how the constituents, that correspond to the nonterminals
(or preterminals) on the right hand side, combine together to form
a construction of the left side nonterminal. For an example of the rules
with rule schemata see the Figure 4.2. The result of the application of
the schema is then subject to the type checking mechanism which safeguards
that the constituents agree in the logical type with the others in
the resulting construction, i.e. that all arguments of a composition have
the types needed by the corresponding function.
In this way, we form the constructions of constituents such are noun
phrases or adverbial phrases up to the level of a clause.
4.2.2.3 Clause Construction
In a rule of the form ‘clause → ...’, the process becomes a little more
complex than to be described in one step only. In such rule, we have identified
the kind of the verb group, i.e. whether it is an attributive or an
132
4. TRANSPARENT INTENSIONAL LOGIC 4.2.2
episodic verb,7
active or passive voice and past, present or future tense.
In groups of so called intersegments we have the candidates for the verb
arguments and free adjuncts in the form of noun phrases, prepositional
noun phrases, other clauses or adverbial phrases. In several successive
steps, we now need to form the construction corresponding to this particular
clause:
1. first, we try to identify the subject (typically Agent) of the clause. In
Czech, we can seek for a noun phrase (including a single adjectival
group, optionally followed by an indeclinable word such as particle,
adverb or interjection) in nominative.
If the subject cannot be determined, we suppose that it is unex-
pressed8
and supply a indefinite subject of the type of individual or
a class of individuals according to the number (singular or plural)
of the verb.
2. after that, we look up the finite form verb in the lexicon where
we obtain all acceptable verb frames of this verb with the corresponding
analyses (that includes the types of the verb arguments,
as well).
3. now, we need to determine the verb arguments in the sentence –
the process is similar to the procedure in the Section 4.2.2.2 above.
In order to reduce the multiplicative extent of the number of participants
to be checked during this process, we run one round of
pruning yet before we start to build the construction – we check all
the intersegments against the available verb frames and first, score
out those that with certainty cannot take part in the verb frame,
and secondly, check all the possibilities (based only on the stated
grammatical categories) of their fitting in place in the verb frame
(e.g. we do not allow two independent verb objects in accusative).
After this, we obtain the possible verb arguments that are then
type checked according to the requirements of the verb.
4. if we have linked in a relative clause or a clause with an inexplicit
subject, we try to supplement it with the subject of the principal
7see the Section 4.3 for description of attributive and episodic verbs.
8In Czech, this is a frequent case – if the sentence subject is a personal pronoun
(I, you, he, she, . . . ) then the subject can be completely left out (unexpressed) since
people are able to recognize it from the ending of a finite verb in the sentence.
133
4.3 4. TRANSPARENT INTENSIONAL LOGIC
clause (i.e. if its verb and the subject agree in number and gender).
Otherwise, we find the clause’s subject as unexpressed.
In this way, we obtain the construction of a clause.
4.2.2.4 Whole Sentence Construction
Eventually, we process the clauses’ constructions according to their conjunction.
The details of their combinations in the sentence building process
can be found in [Hor02].
Thus, following these steps and the guidelines provided in the previous
section, we can accomplish the logical analysis of the whole natural
language sentence.
4.3 Sentence Logical Analysis Using Complex
Valency Frames
In this section we describe the translation of VerbaLex CVFs into constructions
of verb phrases, which form a core of sentence logical analysis.
Valency frames have been built in several projects (VALLEX for
Czech PDT [Žab05] or VerbNet [KKRP06]). Motivation for the VerbaLex
project came from comparing Czech WordNet verb frames with
VALLEX. The main goal of VerbaLex is an automatic processing of verb
phrases exploiting explicit links to Princeton WordNet. The complex valency
frames we are working with can be characterized as data structures
describing predicate-argument structure of a verb which contains the verb
itself and the arguments determined by the verb meaning. The argument
structure also displays the semantic preferences for the arguments. On
the syntactic (surface) level the arguments are most frequently expressed
as noun or pronominal groups in one of the seven cases (in Czech) and
also as prepositional cases or adverbials.
The semantics of the arguments is typically labeled as belonging to a
given semantic role, which represents a general role plus subcategorization
features (or selectional restrictions). Thus valency frames in VerbaLex
include information about:
1. the syntactic (surface) information about the syntactic valencies of
a verb, i.e. what morphological cases (direct and prepositional ones
134
4. TRANSPARENT INTENSIONAL LOGIC 4.3
in highly inflected languages such as Czech) are associated with
(required by) a particular verb,
2. semantic roles that represent the integration of the general labels
with subcategorization features required by the meaning of the verb.
The inventory of the semantic roles is partly inspired by the Top Ontology
and Base Concepts as they have been defined within the EuroWordNet
project [VB+
98]. Thus we work with the general or “large” roles like
AG(ENT), ART(IFACT), SUBS(TANCE), PART, CAUSE, OBJ(ECT)
(natural object), INFO(RMATION), FOOD, GARMENT, VEHICLE
and others. These first level roles are combined with the literals from
Princeton WordNet 2.0 [WN07] where literals represent subcategorization
features allowing us to climb down the hypero/hyponymic trees to
the individual lexical units.9
The verb entries are linked to the Czech and Princeton WordNet 2.0,
i.e. they are organized around the respective lemma in synsets (synonymical
sets) with numbered senses. The inventory of the semantic roles we
work with clearly represents a sort of ontology which tries to cover word
stock of Czech verbs and can be used as a base for a semantic classification
and subclassification of the verbs. The ontologies represent theoretical
constructs designed from the “top” and as such they are not directly
based on the empirical evidence, i.e., corpus data. Thus there is a need
to confront the ontologies and the inventories of the semantic roles that
can be derived from them with the corpus data and see how well they
can correspond to them. For this purpose we are experimenting with the
corpus data obtained from the Word Sketch Engine [KRST04].
The Transparent Intensional Logic comes with a dissociation of significant
verbs into two groups according to the classification of their mean-
ing:
1. by attributive verbs we ascribe qualities or properties to objects.
Attributive verbs are typically expressed by the respective form of
the verb “to be” combined with an expression denoting a property;
examples: “to be red” or “to be mellow” or with a general
substantive like “to be a traitor,” “to be a tree.”
9see the Section 2.1.2.
135
4.3.1 4. TRANSPARENT INTENSIONAL LOGIC
2. episodic verbs, on the other hand, specify actions performed by a
subject. An episodic verb does not describe its subject’s state in any
moment of time, it rather describes an episode of doing something
at the certain time moment (and necessarily some time before that
moment plus the expectation that it will last also in the next few
moments, at least). TIL provides a complex handling of episodic
verbs including the verb tense, aspect (perfective/imperfective) or
active/passive state. All these features are concentrated around the
so called verbal object, the construction of which (i.e., the meaning
of a particular verb phrase) is the application of (the construction
of) the verb to (the constructions of) the verb arguments.
Since the analysis of attributive verbs is usually quite simple, we will concentrate
in the following text on examples of selected episodic verbs from
VerbaLex and their logical analysis using the complex valency frames.
The TIL type of episodic verbal objects is (o(oπ)(oπ))ω , where π is
the type of propositions (oτω). See [Tic80, Hor02, pp. 64-73] for detailed
explanation. Our analysis is driven by a linguistic (syntactic) context
that signals the semantic fact that there is always a function involved
here, so that we have to ascribe types to verb arguments.
4.3.1 Examples of Logical Analysis
In this section, we have chosen 7 verbs with their verb frames from VerbaLex
and we will use them as examples of the algorithm for determining
the verb type in the TIL logical analysis procedure.
dát (give)
dát:2 / dávat:2 / darovat:1 / věnovat:1 (give:8, gift:2, present:7)
give as a present; make a gift of
–frame: DON<organization:1>obl
what nom VERBobl
OBJ<object:1>obl
what acc BEN<person:1>obl
to whom dat
–example: firma věnovala zaměstnancům nová auta (a company
gave new cars to the employees)
–use: prim
The verb arguments in this frame are: who, to whom, what (all obligatory)
with (at least) two options: a) to whom is an individual, b) to
136
4. TRANSPARENT INTENSIONAL LOGIC 4.3.1
whom is a class of individuals. The respective verb types are ad a):
((o(oπ)(oπ))ω ιιι), ad b): ((o(oπ)(oπ))ω ι(oι)ι).
For example, take to whom = “to the employees of a given institution.”
Then “to be an employee of the institution XY” is a property,
say Z/(oι)τω. So “The company gave to the employees of XY. . . ,” not
taking into account grammatical tenses and omitting trivializations we
get λwλt[GivewtXY Zwt . . . ].10
With this example, we can show that CVFs are used not only for
determining the verbal object type, but also for stating additional prerequisities
(necessary conditions) for the sentence constituents. The full
analysis using the verb frame above thus contains, except the verb phrase
part, the conditions saying that “X gives Y to Z ∧ organization(X) ∧
object(Y) ∧ person(Z).” The predicates organization, object and person
here represent the properties denoted by the corresponding terms in the
WordNet hypero-hyponymic hierarchy.
dát:15 / dávat:15 / nabídnout:3 / nabízet:3 (give:37)
offer in good faith
–frame: AG<person:1>obl
who nom VERBobl
ABS<abstraction:1>obl
what acc
REC<person:1>obl
to whom dat
–example: dal jí své slovo (he gave her his word)
–example: nabídl jí své srdce (he offered her his heart)
–use: fig
Here we have an idiom (“to give word”, i.e. “to promise”), which corresponds
to an (episodic) relation between two individuals. Therefore the
type of the verb is ((o(oπ)(oπ))ω ιι), the second ι corresponds to to whom.
bránit (prevent)
bránit:1 / zabránit:2 / zabraňovat:2 / zamezit:2 / zamezovat:2
(prevent:2, keep:4)
prevent from doing something or being in a certain state
–frame: AG<person:1>obl
who nom VERBobl
PAT<person:1>obl
to whom dat
ACT<act:2>obl
inf
10XY has the type ι here, being a collective rather than a class.
137
4.3.1 4. TRANSPARENT INTENSIONAL LOGIC
–example: zabránila mu uhodit syna (she prevented him
from hitting the son)
–use: prim
bránit:1 / zabránit:2 / zabraňovat:2 / zamezit:2 / zamezovat:2
(prevent:2, keep:4)
prevent from doing something or being in a certain state
–frame: AG<institution:1>obl
what nom VERBobl
PAT<person:1>obl
to whom dat ACT<act:2>
opt
in what loc
–example: policie mu zabránila v cestě do zahraničí (police
prevented him from travelling abroad)
–use: prim
Here, arguments of the verb correspond to the phrases who, to whom, in
(from). The third argument has the type of an activity given, of course,
by an episodic verb “hit the son,” “travel abroad” (the substantive form
travelling abroad can be construed as that activity). The type of the verb
is ((o(oπ)(oπ))ω ιι((o(oπ)(oπ))ω )).
říct (say)
říct:1 / říkat:1 / říci:1 / pravit:1 (say:6)
utter aloud
–frame: AG<person:1>obl
who nom VERBobl
COM<speech act:1>obl
what acc,that,dsp
ADR<person:1>
opt
to whom dat
–example: říct kolegovi dobrý den (say hello to a colleague)
–example: řekl, že to platí (he said that it holds)
–example: pravil: “Dobrý den” (he said: “Good day”)
–use: prim
The case questions for the corresponding arguments of the verb “říct” are
a) who, what1, b) who, what2, c) who, to whom, what1, and d) who, to
whom, what2. Examples of instantiated sentences can be a) “Charles says
‘Hello’,” b) “Charles says that he is ill,” c) “Charles says to his colleague
‘Hello’,” or d) “Charles says to his colleague that he is ill.”
The quotation context (ad a), c)) is normally impossible to type.
Unless we want to go into some deep analyses we can ascribe to any
138
4. TRANSPARENT INTENSIONAL LOGIC 4.3.1
quoted expression the type of individual. The relation to and unquoted
subordinate clause is analysed as a general construction of type ∗n. The
resulting types of verbs are then
a) ((o(oπ)(oπ))ω ιι),
b) ((o(oπ)(oπ))ω ι∗n),
c) ((o(oπ)(oπ))ω ιιι),
d) ((o(oπ)(oπ))ω ιι∗n).
brečet1 (cry1) because of something, for something
brečet:1 / plakat:1 (cry:2, weep:1)
shed tears because of sadness, rage, or pain
–frame: AG<person:1>obl
who nom VERBobl
CAUSE<cause:4>obl
due+to+what dat,over+what ins,
for+what acc
–example: brečela kvůli zničeným šatům (she cried for spoiled
clothes)
–example: plakal nad svou chudobou (he cried over his poverty)
–example: plakal pro své hříchy (he cried for his sins)
–use: prim
brečet2 (cry2) for somebody
brečet:1 / plakat:1 (cry:2, weep:1)
shed tears because of sadness, rage, or pain
–frame: AG<person:1>obl
who nom VERBobl
ENT<person:1>obl
for+whom acc
–example: plakala pro milého (she cried for her boy)
–use: prim
If I “cry because of, for” etc., then the role of causing is played by
this because of. Crying is an episodic verb, whereas because of is a relation
between propositions, often between events. We have therefore
because of/(oππ)τω, where the first π (= oτω) belongs to the proposition
denoted, e.g., by “clothes have been spoiled” or that “the respective individual
is poor, sinful” etc., and the second π to the proposition that
the respective individual cries.
In case of “to cry for somebody” the respective type is again a relation
((o(oπ)(oπ))ω ιι), although this for hides some cause, which is, however,
not mentioned.
139
4.3.1 4. TRANSPARENT INTENSIONAL LOGIC
With this verb, we will describe the analysis of verb entailment handling
in TIL. If we analyse a general case of the above mentioned meanings
of cry (cry1 – because of something, cry2 – for somebody) simply
“to cry” (cry3, “He cries all the time”), this verb type is a verbal object
without arguments, (o(oπ)(oπ))ω . In addition to this the following rule
holds: If X cries because of . . . or X cries for . . . , then X cries. In this
way the semantic dependence between the three cases of crying is given;
otherwise we would not be able to detect this connection, e.g. between
cry1 and cry2.
absolvovat (undergo)
absolvovat:2 / prožít:1 / prožívat:1 (experience:1, undergo:2,
see:21, go through:1)
go or live through
–frame: AG<person:1>obl
who nom VERBobl
EVEN<experience:3>obl
what acc
LOC<location:1>
opt
in what loc
–example: absolvoval vyšetření na psychiatrické klinice (he went
through investigation in a psychiatric clinic)
–use: prim
In general “absolvovat něco (undergo something)” is an episodic relation
to an event (type π),11
so the verb type is ((o(oπ)(oπ))ω ιπ). In some
cases we may also use a relation to an episode (specific class of events,
type (oπ)), then the type is ((o(oπ)(oπ))ω ι(oπ)), and “investigation in a
clinic” has to be defined as a sequence of events.
akceptovat (accept)
akceptovat:3 / přijmout:6 / přijímat:6 (accept:4)
react favorably to; consider right and proper
–frame: AG<person:1|social group:1>obl
who nom VERBobl
STATE<state:4>|EVEN<event:1>|INFO<info:1>obl
what acc
–example: akceptujeme jeho povahu (we accept his character)
–example: lidé přijali nový zákon s nadšením (people accepted
new law with enthusiasm)
–use: prim
11see [Tic80, Hor02, p. 65].
140
4. TRANSPARENT INTENSIONAL LOGIC 4.3.1
We can “accept” nearly anything. Here we meet the problem of typetheoretical
polymorphism, which is handled here as a polymorphic type
scheme ((o(oπ)(oπ))ω ια), for an arbitrary type α. A quintessence of such
a polymorphism: “think on (about)” – one can think of an object of any
kind.
učit (teach)
naučit:1 / učit:1 / poučit:1 / poučovat:1 / vyučovat:1 (teach:1,
learn:5, instruct:1)
impart skills or knowledge to
–frame: AG<person:1>obl
who nom VERBobl
PAT<person:1>
opt
whom acc
KNOW<subject:3>obl
what acc,to what dat
–example: naučil dítě abecedu (he educated a children in the
alphabet)
–example: učí studenty matematiku (he/she teaches
mathematics for students)
–example: vyučuje dějepisu (he/she teaches history)
–use: prim
If understood as in “What does (s)he live off? (S)he teaches.” it is
the case of cry3 (see above). “To teach” understood as in “He teaches
history, maths”, etc., the analysis depends on which type is given to
the school subjects, disciplines. One possibility is to analyse them as
properties of a set of propositions, (o(oπ))τω. Then to teach receives the
type ((o(oπ)(oπ))ω ι(o(oπ))τω). If “teaches alphabet” is the case then we
have to decide what we mean by “alphabet.” Here the point is “to teach
(learn) to associate symbols and sounds (phonemes?),” so the respective
type of alphabet is (αβ), where α is the type of symbols, β the type of
sounds. In the analysis of “to educate somebody in something” the verb
takes an individual as its additional argument: ((o(oπ)(oπ))ω ιια), where
α is the type of the discipline.
In all the examples, we have displayed the relations between the twolevel
semantic roles used in the VerbaLex verb frames and the resulting
logical analysis types of the verbal object as the main part of the clause’s
logical construction. The algorithmization of this procedure uses a list
of all roles used in the lexicon (there are about 200 roles used) with the
141
4.4 4. TRANSPARENT INTENSIONAL LOGIC
corresponding (ambiguous) logical types of the constituents. In this way
we can form a basic skeleton of the automatic translation of text to logical
constructions.
4.4 TIL Knowledge Base Representation
We can look at the knowledge base (KB) design as consisting of the ontology
and the description of the way how the logical inference is performed.
The ontology is used as a specification of all concepts, that the inferring
machine knows about. Only those concepts that are committed to the
ontology can ever be thought of as being processed by the machine when
used in a question-answering or a dialogue system. In our case the ontology
is well defined by means of TIL. All objects that are stored in the
knowledge base are derived directly from input constructions that reflect
the meaning of facts that the user had implanted to KB. Every object
has an equivalent type in the epistemic framework of TIL.
The system knowledge base design is a form of a semantic network
consisting of declarative memory for storing facts and procedural memory
for storing inference rules (a similar approach is used, e.g., in the Soar
system [RLN93]). Facts in the declarative memory are strictly time-dependent,
which means that each fact is bounded to the time moment,
at which the fact was stored.
The KB structure can be illustrated by the example representing the
facts that arise from the input sentence
John told a story to his daughter yesterday. (4.2)
The equivalent TIL construction is stated in the Figure 4.3.
When storing the facts in KB, the existentially quantified variables
are replaced with newly allocated constants of the appropriate type (by
the process of skolemization) or linked to already allocated objects.
The part of KB that represents our construction consists of 15 simple
type objects (objects without subconstructions), 2 variables (w1 and t1)
and 13 structured terms (K1, . . . , K13). The whole semantic network that
represents this sentence is depicted in the Figure 4.3. The objects I1 and
C1 are the constants that replaced the skolemized variables x and z. Two
other constant objects t0 and wTim refer to the moment of the utterance
of (4.2) and the reference world of the TIL inference machine. Those
142
4. TRANSPARENT INTENSIONAL LOGIC 4.4
P
Onc
Does
Perf
tell
wTim
C1I1
John
Yd
t0
t1
w1
Of
daughter
⊂
story
K1
K4
λ
K13
K12
K11
K9
K10
K2
K3
λ
K7
K5
K6
K8
λwλt
2
6
6
4Pt
"
Oncwλw1λt1(∃x)(∃z)
»
daughterw1t1
x ∧
h
Ofw1t1 x John
i
∧
∧
h
storyw1t1
z
i
∧
h
Doesw1t1 John
ˆ
Perfw1 [tell x z]w1
˜i–#
[Yd t]
3
7
7
5
(4.1)
Figure 4.3: The part of KB with the construction of the sentence (4.2).
143
4.4.1 4. TRANSPARENT INTENSIONAL LOGIC
constants were put into the construction (4.1) in place of the abstracted
variables w and t.
Such format of the knowledge base properly reflects the reuse property
of constructions and subconstructions. This property is necessary
for large knowledge databases, since they share the space of identical
constructions, and on the other hand, it helps to identify all the constructions
working over one topic (expressed by a subconstruction) in a
very straightforward way.
4.4.1 Knowledge Base Implementation – the Dolphin
System
Dolphin is a part of a complex project employing TIL to develop a computer
system with data understanding ability. Such system will be able
not only to store natural language structures but also to mine new facts
logically resulting from input and the data stored before. One of the
final benefits of the whole system will be no need for special humanmachine
communication language – the user will just ask and Dolphin
will answer what can be inferred from the knowledge base to the question.
The Dolphin system is thus an experimental implementation of the
TIL knowledge base developed by Andrej Gardoň in the MU NLP Centre
under the supervision of Aleš Horák.
The Dolphin system is designed to process the output of the syntactic
parser synt (described in the Chapter 3), which is able to produce
syntactic trees as well as logical representations of input sentences in the
form of TIL constructions. In Dolphin, the constructions are parsed and
stored in the knowledge base (KB), which takes the form of a semantic
network. Currently, the Dolphin system stores labels of natural language
objects described in the Czech language. However, the actual chosen
language does not limit the functionality of the system. The Czech language
is used in Dolphin because the transcription of natural language
sentences to TIL constructions is handled by the Normal Translation Algorithm
(NTA) system [Hor02] that presently provides only the Czech
transcription ability.
When a new input is to be stored in the database, it has to pass
through several parts of the system. The Figure 4.4 demonstrates the
process of storing new data to Dolphin.
In the first phase of storing a new fact saying “Lemon is yellow,” the
144
4. TRANSPARENT INTENSIONAL LOGIC 4.4.1
sentence
NTS
Dolphin
sentence
NTA lemon/(oι)τω, yellow/(oι)τω
True : ∀x[lemonwTimt0
x ⊃ yellowwTimt0
x]
Lemon is yellow.
learned
OK
Figure 4.4: The phases of data flow in Dolphin.
sentence is translated by NTA to the TIL construction of the sentence.
Dolphin then parses the construction and stores it in the knowledge base.
At the level of propositions, Dolphin interprets logical operators and divides
the sentence to several basic facts. The composed construction
is stored as a connection among sub-constructions. The last symbol of
the sentence (., ?, !) indicates the mode of the database operation
(storing/answering/ordering). NTS stands for the Normal Translation
Synthesizer and its role is to produce a sentence from TIL descriptions
provided by the output of Dolphin.
4.4.1.1 The Dolphin Storage
As we mentioned before, the chosen input natural language does not
influence the functionality of the Dolphin KB. On the other hand it is
very important in searching for objects. The Dolphin KB thus stores the
data in two layers: the language layer and the object layer.
The basic Dolphin idea is a separation of the language layer from
the logical (object) layer. Human words are just a way how to describe
some objects. There are many languages but all of them work over the
145
4.4.1 4. TRANSPARENT INTENSIONAL LOGIC
same set of objects. The Dolphin input processing lies in transforming
its teacher’s words to the KB objects mentioned above. Here synt plays
its role as it produces TIL transcription of a sentence. E.g., Dolphin
after obtaining apple on its input takes its TIL description and creates
a new object in its object layer – let it be the object #1. Next time we
mention this particular apple, Dolphin looks up the object #1 in its KB
and whatever is told about it is linked to this object.
The Language Layer This layer provides the language encoding ability
of the Dolphin system allowing the rest of the system to be completely
language independent.
The implementation of the language layer (LL) has the form of multilingual
dictionary providing ILIs (inter lingual indexes) for the object
layer. Each LL record contains an object name expressed by words in
different languages.
The Object Layer This layer is quite complex so only basic features
are mentioned, see [Gar07] for details. Each TIL object is given a type
from the type hierarchy built over four basic types – ι (individuals as
object labels), o (truth values, True and False), τ (time moments and
real numbers) and ω (possible worlds).
One of the essential KB objects for inferring the object properties are
the objects of types including ι in it. Thus those objects are stored in
separate files containing this information:
• the type of the object
• the connection to the language layer, provides naming of objects
using words of natural language
• membership – many objects form classes or relations. To provide
fast way of searching for data, it is essential to know in which
classes/relations an object has a membership
• members – only for objects with composed type to identify simple
objects that form them
Objects of type τ and ω are currently just stored in the database without
any special handling.
146
4. TRANSPARENT INTENSIONAL LOGIC 4.4.1
4.4.1.2 The Input Sentence Processing
Let us take an example input sentence (in Czech, as this is the current
input language of synt):
Jablko je červené. (An apple is red)
and its corresponding TIL transcription:
λw1λt2(∃i3)([0
jablko-0w1t2
, i3] ∧ [0
červený-2w1t2
, i3]) . . . π
At first, variables are identified and single applications are isolated. Variables
that are not covered by λ-abstraction are replaced by new constants
named after the variables. The λ-abstracted variables are currently handled
in a simplified way – when the λ-abstraction is used without quantification,
the abstracted variables usually go over the ω (possible worlds)
or τ (time moments) types. In case of an ω-variable, the Dolphin’s world
wDolphin is assigned to it. All time variables (of type τ) are, in the
current version, replaced by an object representing a “general time independent
object,” so as we can see the current version does not support
time processing at all. This feature will be the main goal of our future
work. Other quantified variables are initialized according to the system
running mode. There are two running modes currently available – the
learning mode and the question answering mode.
The Learning Mode processes and stores new facts and it is activated
by a full stop mark at the end of a sentence. In this mode, each uninitialized
variable is replaced by a new object named after the variable. If
we want to add new facts about an object previously mentioned in the
conversation with Dolphin, we have to stress this with the demonstrative
pronoun (ten, ta, to – in Czech this corresponds to the definite article in
English). The object is then marked with an exclamation mark (!) and
is looked up in the knowledge base and the variable is replaced by the
stored object. For example if we want to add the fact that
To jablko je červené. (The apple is red.)
the apple is analysed as 0
!jablko-0/(oi)τω and then found as the previously
stored object #23 and the construction would thus contain object
0
#23 instead of the i3 variable displayed above. Currently, the system
147
4.4.1 4. TRANSPARENT INTENSIONAL LOGIC
supports only the existential quantification since the universal quantification
defines new inference rules and the complex inference has not been
implemented so far.
We may mentioned the way, how possible fact conflicts are handled –
if the input fact assigns something to an existing KB object and the fact
is in conflict with the KB content then an error message is raised. For
demonstration let us have a sentence
The apple is red.
stored in KB. Now we would like to store the sentence
The apple is not red.
This raises an error message and the second sentence is not stored. Today
there is no tool for saying to Dolphin:
A fact in KB is wrong, I have the correct one.
so whatever is stored in KB will remain unchanged until a reset of the
whole database.
The Question Answering Mode works similarly to the learning
mode but there is no unification of free variables with new objects. Instead,
the first application containing an uninitialized variable suggests
a way how to unify this variable with a particular value. If we take our
example sentence as a question
Je nějaké jablko červené? (Is there any red apple?)
first application containing free variable i3 (variables w1 and t2 are initialized
as described above) will be
[0
jablko-0w1t2
, i3]
Thanks to this application, i3 is unified with an object that is stored in
Dolphin and is an Apple (is linked to the class construction of apples).
The following application (saying that the object is red) posts a second
requirement to check. If the object unified with i3 is not Red (there
is no relation between the unified object and the class of red objects),
the system returns to the state when i3 was initialized and tries another
148
4. TRANSPARENT INTENSIONAL LOGIC 4.4.1
possibility. If all possible ways were checked and there is no object that
is Apple and Red at the same time, the answer NO is returned. In case
all the requirements are fulfilled, the answer YES is returned with the
selected i3 value. Again there is no universal quantification feature yet,
but simple questions such as
Jaké červené objekty znáš? (What red objects do you know?)
can be processed correctly.
4.4.1.3 How the Objects are Stored
Objects are essential elements of logical analysis and the TIL construction
processing. In the Dolphin knowledge base, all data are objects
related with each other. It does not matter if we store an individual or
a function in KB, we are still working with one object. For example,
the class of apples (Apple) is represented as the characteristic function
of the class, i.e. as a (world and time dependent) function which returns
an o-object (boolean true/false) for each ι-object (an individual) given
as its argument. Each such ι-object is stored separately from Apple, of
course.
There is an information that Apple and the corresponding ι-objects
are in relation and this information is shared among all participants.
Each object is represented as a separate file. In the future, a specialized
file system technology can be developed or archiving methods can be
incorporated in the storage process without the database functionality
limitation or reimplementation.
4.4.1.4 A Complex Example
Let us take three example sentences:
1. Toto je jablko. (This is an apple.)
2. Tato kostka je červená. (This cube is red.)
3. To jablko je červené. (The apple is red.)
and their TIL transcriptions provided as inputs to Dolphin, step by step:
λw1λt2([0
jablko-0w1t2
, T oto]) . . . π
149
4.4.1 4. TRANSPARENT INTENSIONAL LOGIC
The word toto (this) makes the resulting construction an open construction
(with the free variable Toto) which needs to receive so called pragmatic
anchor.12
This input creates a new object in the knowledge base
and Dolphin prints the assigned object number:
> Toto je jablko. (This is an apple.)
> stored as object 6.
The second sentence
> Tato kostka je červená. (This cube is red.)
with the corresponding TIL construction
λw1λt2([0
!kostka-0w1t2 , i3] ∧ [0
červený-2w1t2
, i3]) . . . π
where the variable w1 is unified with object #2 (representing the Dolphin’s
world), the variable t2 is unified with object #3 (representing the
General time object) and the variable i3 is now initialized with a new
object of type ι (let it be #7), since no previously mentioned object from
the class kostka was found.
Each trivialization asks the language layer whether it knows the word.
If the word is not found, it is stored and a new object is created with a
connection to this word.
So ∧ (AND) in our transcription causes that a look-up in the language
layer is performed and the object #10 is returned (we suppose that AND
was previously stored).
Applications are represented as relations among objects that are participating
in the application. Thus in Dolphin, the partial application
[0
červený-2/(oi)τωw1]
is represented as a relation between červený and the object of the variable
w1. The information about the relation is stored both in the červený
object and w1 object and of course in the final object – the result of this
application is a new (oι)τ-object (a chronology of the class of objects
which are červený/red) which is than applied on the general time object.
The final object of the application
[0
červený-2w1t2
, i3] . . . o
12see [Hor02, the Section 5.2.4] or [Mat98, the Section 7.1]
150
4. TRANSPARENT INTENSIONAL LOGIC 4.4.1
Figure 4.5: The semantic network for the 3 example sentences.
is then created as an o-object (a truth value) and in the learning mode
the object receives the value of True.
The third example sentence is stored similarly as the second one with
the i3 variable definition difference.
> To jablko je červené. (The apple is red.)
TIL transcription
λw1λt2([0
!jablko-0w1t2
, 0
#6] ∧ [0
červený-2w1t2
, 0
#6]) . . . π
151
4.4.1 4. TRANSPARENT INTENSIONAL LOGIC
The semantic network for three sentences is displayed in the Figure 4.5.
Note that there are many objects with the value of True in the network.
This is because each application leads to a unique object and if this
object is not yet in KB, it is created. Thanks to this, the λ-abstraction
can be done in a straightforward way. It is worth saying that negation
is a special function over the o-object that does not create a new object
but it replaces the existing True value with False. If we ask the question
Je to jabko červené? (Is the apple red?)
TIL transcription has the form of a match:
x . . . o : ([0
!jablko-0wDolphintnow
, 0
#6] ∧
∧ [0
červený-2wDolphintnow
, 0
#6]) . . . o
Basically it asks whether the previously mentioned object #6 is in relation
with Apple and Red at the same time. To answer this, the system obtains
the result of the applications on the right side of the match and checks
whether the final object of the right side is True. In this case, the answer
is YES. But what happens if we have the question
Je to jablko zelené? (Is the apple green?)
Since we are in the question answering mode, the basic trivialization of
zelený (green) will fail as we do not have such object in KB yet. The
system does not follow the predicate logic closed world assumption, thus
the answer is “I DO NOT KNOW” instead of “NO.” The question
Které červené objekty znáš? (What red objects do you know?)
is again analysed as a match
s . . . (oi) : 0
červený-1wDolphintnow
. . . (oi)
Here the system evaluates all possible objects which are stored in KB as
related to the class červený (applied on the Dolphin’s world and the general
time object) obtaining the class of all red objects known to Dolphin.
Thanks to the system design, this operation consists in fast searching
through the semantic network. As can be seen in the Figure 4.5, the
Red object applied on the Dolphin’s world and the General time object
((oi)-object with (9,2) 3 label) has connections to the True value. Now
152
4. TRANSPARENT INTENSIONAL LOGIC 4.5
(simplified) it is enough to look what object is at the end of the connection
that runs out from the place of the True value where the connection
from (9, 2) 3 object ended. In this way we find out that objects #6
and #7 are red.
4.4.1.5 The Development of the Dolphin Inference Machine
Nowadays, the Dolphin Inference Machine (DIM) is at the start of its
development. Currently, DIM is equipped only with basic modus ponens
like inference mechanism and is able to answer question from conversation:
“Every boy is a child. Peter is a boy. Is Peter a child?”
The Dolphin system is able to parse and store the synt output in the
form of TIL constructions, to check the consistence of the database and
to answer simple questions. The final aim of the Dolphin development is
the full support of working with possible worlds and time moments and
the provision of a complex inference tool.
4.5 Experiment of Using TIL in a Simulation
System Easel
Easel [Fis99] is a simulation system designed for gathering, processing
and display of information about various active entities of the simulated
world, about their interactions, and about their collective global effects.
For the description of the entities in the simulated processes, Easel
uses a property-based type system. In the system a type provides a
(complete) description of a template of an object. Any entity is then an
instantiation of such template. The type can also be thought of as a class
of all objects that comply with the set of properties stated in the type
template.
Together with the inheritance of type properties and arrangement
of types in one hierarchy (with the root type all), Easel language resembles
common object-oriented (OO) programing languages and formalism.
However, as an extension to the usual OO classes, Easel emphasizes the
need for type manipulations with symbolic techniques. The definitions
of subtypes are described with adjective modifiers similarly as in natural
language expressions:
153
4.5.1 4. TRANSPARENT INTENSIONAL LOGIC
example(): action is
fruit: type;
round(any): type;
green(any): type;
apple: type is round fruit;
golden_delicious: type is green apple;
my_apple: golden_delicious;
confirm my_apple isa fruit;
example();
4.5.1 Using TIL and Easel in Applications
The specifications of Easel types is suitable for limited simulations with
strictly specified bounds of the simulated world. However, in case of
complicated descriptions of such extent that reaches the complexness of
the real-world situations, the descriptiveness of the types in Easel would
run against difficulties due to their underspecification.
The possible combination of both, TIL types and Easel types, in one
application seems to be a promising completion of the type specification
techniques. TIL allows to place no limits to the expressiveness of the
descriptive language used to describe the types and subtypes used in
the simulation and (possibly) can make use of any properties from the
intensional base of natural language.
The confirmative claims that are currently allowed in the Easel language,
would not need to be closed to browsing the type hierarchy, but,
with the use of the inference mechanism in TIL, it could analyze any
consistent proposition about the type. The above stated example could
look as follows:
example(): action is
# the primary properties
Fruit is a class of individuals.
To be round is a property of individuals.
To be green is a property of individuals.
# the definitions and instantiations
Every apple is a round fruit.
Golden Delicious is a green apple.
154
4. TRANSPARENT INTENSIONAL LOGIC 4.5.1
The variety of my apple is Golden Delicious;
# statements
If I eat a fruit, I am healthy.
# claims to be confirmed
confirm that my apple is a fruit;
confirm that if I eat my apple, I am healthy;
example();
The form of the description of the situation, i.e. the type definitions, can
be written in natural language statements about the discussed elements
and their properties. An automatic parsing system can translate the
specifications into the following objects and propositions:
example(): action is
# the primary properties
fruit . . . (oι)τω
round . . . (oι)τω
green . . . (oι)τω
# the definitions and instantiations
(∀t)(∀x)applewTimtx ⊃ (fruitwTimtx ∧ roundwTimtx)
(∀t)(∀x)0
(Golden Delicious)wTimtx ⊃ (applewTimtx ∧ greenwTimtx)
(∃my apple)0
(Golden Delicious)wTimt0
my apple
# statements
(∀t)(∀x)
h`
fruitwTimtx ∧ DoeswTimtI [PerfwTim [eat x]wTim ]
´
⊃
⊃ healthywTimtI
i
# claims to be confirmed
confirm (find a match)...
x : fruitwTimt0
my apple
confirm (find a match)...
x : DoeswTimt0
I [PerfwTim [eat my apple]wTim ]) ⊃ healthywTimt0
I
example();
In the case of argumentations about instances of a type, like my apple
here, we use the current reference time t0 and the actual reference world
wTim, which are related to the working environment of the processing
system.
The mechanism of analyzing questions and inferring derivations from
other facts is described in more detail e.g. in [Tic82, Hor02]. Finding a
match means a request to infer (from the underlying knowledge base) the
155
4.6 4. TRANSPARENT INTENSIONAL LOGIC
value x of the specified TIL type, i.e. the truth-value o of the extensified
proposition on the right hand side of the match, here.
Even if the above stated examples do not use more power, than a
temporal first order logic system would provide, the TIL theory ensures
correct analysis of much more complicated expressions, like intensional
roles or belief sentences.
A straightforward reuse of the TIL analysis and TIL inference system
is, however, not so easy as a simple incorporation of some other, fully
described algorithm. TIL analysis needs to work with something as complex
as the natural language and the ambitions of TIL do not allow to
make simplifications of the matter.
4.6 Long Way to Full Natural Language Se-
mantics
In this chapter, we have introduced the main ideas and basic structures
of Transparent Intensional Logic. We have also argued for the TIL constructions
being the meaning bearer for any natural language expressions.
The part of the Normal Translation Algorithm for logical analysis
of a natural language sentence that is responsible for building a single
clause construction was explicated here. We have concentrated on the
description of the lexicons needed for analysis of the lexical items and for
analysis of particular verb frames. Several examples of existing lexicon
entries have been displayed. We have also proposed and specified a way
of combining the descriptions of lexical entries in the dictionary with the
respective information obtained from lexical semantic resource, namely
WordNet.
The text presented a first outline of comparison and integration of
the two approaches, namely logical and linguistic, to the semantics of
verbs in a natural language. We are aware that this work is still in a
great progress and the results so presented are preliminary. Still, we are
convinced that the research task we aim at is a relevant contribution to
the semantics of natural language.
We have shown that pursuing such a research is reasonable and comes
up with a new viewpoint to the meaning of verbs. In this way we extend
our knowledge in the important way. Actually, we are dealing with two
deep levels of the meaning description and a question may be asked which
156
4. TRANSPARENT INTENSIONAL LOGIC 4.6
one is deeper and better. Our answer is, do not contrast the two levels,
and make use of both of them. In this way we believe to integrate them
into one compact whole and perhaps obtain a unique data structure. The
results of the presented research can be immediately applied in the area
of knowledge representation and in the long-term Normal Translation
System project that is being prepared. We have not tackled the other
deep descriptions, such as the method that exploits the tectogrammatical
level as it is presently applied in PDT [Haj04a]. This, obviously, is a topic
of another text.
We have briefly presented basic features of the Dolphin system for
handling structured knowledge of TIL with the question answering capa-
bilities.
As a complement to the extended type system of TIL, we have presented
an experiment with the Easel language with its property-based
types. We have formulated the assets of using the two approaches in one
application and showed the need for working implementations of TIL
natural language analyzer and TIL inference system based on the work
of TIL’s author and other researches.
157
4.6 4. TRANSPARENT INTENSIONAL LOGIC
158
Chapter 5
Application in Dialogues
– the Electrical Power
Systems Simulation
Emergent systems and intelligent agents are in the focus of intense attention
in many areas of computer science and artificial intelligence. Their
main advantage can be seen in the automated adaptation to changing environment
and in maintainable control of such system behaviour. There is
no need to specify global constraints for the system as a whole. The high
degree of autonomous behaviour means that the user can concentrate on
the most relevant areas.
Our research focuses on a future architecture of electrical power system
simulations able to cope with the changing conditions. The electrical
power industry in the Czech Republic and all post-communist Central Europe
countries changed dramatically in the last decade. The changes on
the market and the development of information technology as well as the
power sector call for new methods that are able to cope with the evolving
environment. There is a need for flexible adaptation of the current power
system operations. Optimizations of power networks have to take into
account the market strategy. The maintenance and network planning
must be in a balance with the return of investment for the distribution
companies.
159
5. APPLICATION IN DIALOGUES – EPS SIMULATION
Intelligent agent systems are described in many new artificial-intelligence
monographs (see e.g. [RN03]). The application of intelligent agents
in power systems control and operation is discussed in [Reh03]. Intelligent
autonomous systems are seen as systems that can react intelligently
and flexibly on changing operating conditions and demands from the surrounding
processes. The analysis of electrical networks from the agent
perspective is tackled also in [Aea93].
To enable interconnection of the simulation system with the current
power system knowledge bases (e.g. with actual information about the
reliability of the used components and current failure data), the systems
have to employ standard communication technologies, protocols
and models. There are many de facto standards as well as official international
standards that the system should be complied with. For
example, a communication standards are specified in IEC 61870: Telecontrol
equipment and systems – Part 6: Telecontrol protocols compatible
with ISO and ITU-T recommendations (2002) or IEC 61850: Communications
networks and systems in substations. Another useful standards
for the purposes of data interchange are XML and, especially, web services
standards such as SOAP [SOA03] or newly proposed WSRF (Web Service
Resource Framework [C+
04]). The interfaces of the developed systems
should be compliant with IEC 61968: System Interfaces for Distribution
Management and IEC 61970: Energy Management Systems Application
Programming Interfaces.
The fault-tolerance and quality assurance of the power supply is very
important and constantly evolving science branch which demands intensive
research. Because of the geographic dissipation and big financial cost
of the electrical power systems (EPS) facilities maintenance, the required
reliability cannot be reached simply by redundancy. It is necessary to
search for new low-cost but effective means of the required reliability
assurance.
One of the possible approaches to this problem is regular checking
of the facility operation. Failure data are gathered in failure databases
with a precise description of the failure cause, time, severity and the
description of the failed component – its manufacturer, model, serial
number etc. This effort allows a statistical analysis of the reliability of
the particular facility series, types etc.
This type of the failure database is for the whole Czech republic built
and maintained by the research team for creation and categorization
160
5. APPLICATION IN DIALOGUES – EPS SIMULATION
of failures records for distribution equipment and outages of supply at
all voltage levels in the Technical University of Ostrava in the team of
Stanislav Rusek. Utilities for the analysis of the data in this database
are also developed in Ostrava in the team of Václav Snášel. The aim of
this effort is to develop a new methodology for the power system facility
maintenance based on the condition-centered approach rather then the
contemporary used preventive time-based approach. Power system facilities
are well tested in the process of their design and manufacturing but
there’s a serious lack of testing in the process of their real operation.
The new methodology should be flexible enough to allow fast and
effective control even for facilities that lack measured reliability data.
The Technical University of Ostrava with the help of other universities in
the Czech Republic has made an interdisciplinary team which is focused
on the development of new methodology for the power system facility
condition checks, outage prediction, black-out risk reduction and early
post-fault recovering.
The team at FI MU Brno is focused on the development and usage of
multi-agent systems in power system facility simulation, monitoring and
control.
Enterprises running power systems are making big efforts to make the
reliability assurance process more flexible and adequate to the real conditions
of the power facilities. Simple time-based controls of the facilities
do not take into account the rate between the maintenance costs and the
money loss caused by the failures in the energy delivering process.
A more appropriate way of dealing with the reliability assurance will
be the new methodology of maintenance based on the actual state of the
facility and the amount of the energy delivered (or “potentially undelivered”)
by the facility.
To achieve more reliable power delivering many of the facilities are
backuped. Routine action after failure is to change the active lines topology
or to switch to a backup facility to restore power delivering. Each
failure is precisely recorded in a failure database (SCADA, Supervisory
Control And Data Acquisition [SB95]). The process of the power system
maintenance and management, failure recognition and function restoration
can be optimized at least from the economic point of view. The
problem is somewhat classical: to minimize the overall costs (maintenance,
restoration, undelivered energy) which is a function of the facilities
condition and its features, dynamic network topology and also the
161
5. APPLICATION IN DIALOGUES – EPS SIMULATION
restoration process methodology. The problem is that the function is
not computable in a straightforward way and it is not even clear which
variables are to be measured to figure out the function.
Our work has two main motivations: (1) to support the development
of the condition-oriented facility maintenance methodology by the information
technologies, and (2) to develop a software for simulation of the
various features of the power system facilities. The task is concentrated
mostly on simulation of the electrical energy flows, outages due to facility
failures, and their impact on the overall sum of the non-delivered energy.
In its first stage, the developed system (called Rice) is meant to
be a research tool extending classical tools for reliability assurance like
SCADA databases and statistical tools for failure-data exploration. There
is a narrow connection between the exploration of the data in SCADA
databases and the developed simulator.
The simulation of power networks represents an appeal to the research
thanks to its complexity and also to the potential positive economic aspects
of the outage prevention [Ami01]. The ability to compare the results
of potential situations leading to a network outage and even the ability
to automatically compute the economic loss of the engaged power consumers
are the most welcomed features of the Rice system described in
this chapter.
Rice is a prototype of a software system designed for simulation of
the energy flow and the errors and outages propagation in the power
networks developed at the MU NLP Centre by Miroslav Prýmek under
the supervision of Aleš Horák. The architecture of Rice is based on the
principles of the multi-agent software design [Jen00]. The aim of the work
is to develop a software tool which will be able to simulate complicated
behaviour of the power systems in the dynamically changing environment.
The basic idea of the system design, according to the multi-agent approach,
lies in modelling the whole system “from the bottom upwards.”
Definition of a primitive component behaviour and relations and communication
between them is the main principle of the definition of the whole
system characteristics.
Parts of the project work described in this chapter have been published
in [SPHS05, PH05a, SHKC05, PH05b, HP06c, PH06b, HP06b].
162
5. APPLICATION IN DIALOGUES – EPS SIMULATION 5.1
5.1 Multi-Agent Framework for EPS Simulation
and Monitoring
The first phase of analysis and design of the framework for description and
control of electrical power networks processes was inspired by artificial life
(AL) systems. We have taken the most influencing features of biological
systems, which are difficult for simulation by AL methods. The following
section deals with the criteria, but applied to the simulation of electric
power networks.
Topology Electric power networks (EPN) topology can be multi-path,
i.e. in the oriented graph representing an EPN, there could be several
different paths from node A to node B. But on the other hand,
the topology is quite stable and defined precisely.
As opposed to the biological systems, EPN topology is simpler and
hence easier to simulate – it has no features of randomness, dynamics
or inaccuracy, which cause the most problems in modelling
a living system.
Interaction EPN behaviour is limited to local interactions. This fact
enables solving the problem of simulation by means of the network
of many locally interacting agents, which do not depend on inner
states of each other and act according to their own information and
intentions.
Emergence/genesis Like in biological systems, the complexity of EPN
behaviour (and, hence, the difficulty of its prediction) is expected to
depend on the complexity of the system’s structure, and parallelism
and independence of its elements.
The behaviour of particular basic elements of EPN is predictable,
which makes the simulation straightforward and highly plausible. If
the element’s behaviour can be predicted analytically, we can apply
the “behavioural” approach (see [Ski57]), i.e. we treat an element as
a black box, observe and fix its explicit reactions on various stimuli,
and reproduce the same behaviour with the simulator.
Here we should clearly state two hypotheses, which compose the
basis of our work:
163
5.1 5. APPLICATION IN DIALOGUES – EPS SIMULATION
1. the behaviour of the separate EPN elements is easy to determine,
define and simulate.
2. if we unite separate elements into a network and supply the
sufficient support for their parallel, fast (not only real time)
and adequate connectedness, the whole system will appear,
whose complex behaviour is comparable to the real EPN.
Evolution The goal of an AL simulator is to create a system with “openend
evolution” [Ray07], while the goal of the EPN simulation is to
approximate the real functioning of a real EPN. In this case, we
are not interested in the evolution of the whole system; rather we
focus on the approximation of elements behaviour in reality.
The law of evolution in the form of the evolution algorithm is useful
to the EPN simulation, but with different goal and in different
format. Instead of the competitiveness between separate elements
(“organisms”) we must apply the optimum searching algorithm.
Fitness Function As could be seen from the previous paragraph, in
case of EPN the fitness function plays a role, which differs from
that in AL. For our purposes, we define the fitness function in
such a way, that it reflects the extent of the difference between the
simulation and the real element. An advantage can be seen in the
fact, that the fitness function of particular elements of EPN is easily
measurable, as opposed to biological systems, where we often face
the “hidden” fitness.
The arguments stated above clearly show that methods and models, applied
to biological system simulation, are at least to the same extent
effective for the EPN simulation. This conclusion is based on the fact
that, in case of EPN, the criteria demonstrate lower complexity, sometimes
absence of the elements or features, which cause the most problems
with the predictability and effectiveness of the simulation.
For the implementation of the methods, we have decided in the first
phase to adapt the Kairos system [Prý04]. This tool is modular and is
based on rules of object-oriented programming. Primarily, it was designed
for biological systems simulation. Solving the problem of EPN
modelling, some features of Kairos were modified significantly, as to adapt
to the new goal of research. Kairos is designed for Unix, but it can be
164
5. APPLICATION IN DIALOGUES – EPS SIMULATION 5.1.1
ported to other platforms. Three independent blocks compose the original
version of Kairos, which interact through the TCP/IP; only two of
them are applied for EPN simulation – the functional core and the visualization
tool. Some modifications that were made to adapt Kairos to
the goals of EPN modelling are described in the following paragraphs.
Network nodes The network nodes are static according to the topology
and that is why we identify them with life environment fields
in Kairos. Herewith, some features that are usually unique to organisms
must be used – adaptability, communication, complex behaviour
etc. The network nodes are therefore implemented by synthesis
of features for fields and organisms. The leading wire is also
a network element for the sake of error simulation, of course.
Topology The Kairos system has no limitations to the topology of the
simulated environment, since Kairos is based entirely on the local
interactions between environment fields. We exploit this feature
and, instead of two dimensional rectangular grid that is usual in
the life system simulators, we use a structured network appropriate
for modelling real electric power networks.
Used primitives The Kairos system provides several specialized primitives,
out of which we use mostly the current and energy transformation
primitives. The only change that is needed for the sake of
effectiveness is the definition of energy as a relative value. With
this, we do not need to model a constant energy current with a
constant data flow, but only with a single information about the
starting values of the current.
Visualization The visualization subsystem for EPS processes has been
completely redesigned, the energy flows visualization techniques do
not share many common features with artificial life simulations.
Here, the subsystem design is based on existing concepts and implementations
stated in [Reh03].
5.1.1 The Rice System Architecture
Further development of the EPS simulation evolved in a complete new
system called Rice. The Rice software framework is an experimental implementation
of a power system simulator, which implements all of the
165
5.1.1 5. APPLICATION IN DIALOGUES – EPS SIMULATION
above-described features needed for developing the new reliability assurance
methodology. It is based on these main principles, described further:
(1) it is a multi-agent system, (2) it can be used to study behaviour of the
dynamically changing environment, (3) as to practical implementation,
it is decentralized, modular and open. This is necessary especially for use
in geographically dislocated areas (e.g. online facilities monitoring).
The design of the Rice multi-agent framework is based on the following
basic requirements:
Decentralization Because of the power systems nature (especially their
geographic dissipation), the system allows decentralized distributed
operation on many computers interconnected with the standard
type of computer network (LAN, WAN, Internet). This requirement
is even more important for the process of real-time on-site
monitoring of the facilities.
Platform independence The application environment is very heterogeneous.
It is absolutely necessary to use technologies independent
on hardware platform and operating system. The framework should
also support wide range of programming languages.
Performance scaling Hardware requirements of the simulation and the
range of future applications is hardly predictable. The framework
therefore supports easy performance scaling by extension of hardware
and software dedicated to it. This extension should not require
large changes to the existent architecture and implementation.
Modularity and extensibility The framework is flexible, open and
extensible for addition of new features and modules which appear to
be needed during the process of dissemination and testing. In such
case, changes to an existing code should be as minimal as possible.
Open standards We must take into account a possible future need to
integrate the proposed system with the contemporary software and
hardware solutions of the power system facility vendors. To make
this integration easier, the system is based on standard technologies
– with open specifications and implementations.
Security Communication between system modules usually goes through
untrusted channels. Privacy and authenticity of the particular com-
166
5. APPLICATION IN DIALOGUES – EPS SIMULATION 5.1.1
munication acts must be still guaranteed on the level common in
the industry solutions.
Low-cost practical application The framework is to be based on commonly
used technologies. Only then we can assume an easy and
low-cost practical application in the production environment. If
any usage of non-common technologies will prove to be necessary,
it should not appear in many parts of the system, especially not in
the user-oriented parts of the system.
After the specification of all the system requirements, the decision was
made to implement the system as a multi-agent system based on the
standard technologies commonly used in this area – a combination of
two communication protocols, CORBA and KQML. CORBA standard
has freely available specification and there are many implementations –
proprietary ones (e.g. Visibroker, PeerLogic, IONA) and even free ones
(e.g. OmniOrb, Orbit, Mico). In our prototype implementation, we use
OmniOrb (in Python) and Sun ORB which is part of the Java SDK. An
indispensable advantage of the CORBA standard is that all ORB implementations
conforming to the specification would integrate well within
the system.
Usage of the CORBA-KQML solution was presented in several projects,
e.g. [DTH99], [COB04], [BSS98]. The multi-agent approach has
been used even in the area of the power systems monitoring and control
(e.g. [NS04], [MMH03], [CVJ92]). The Rice implementation is, however,
the first system that combines these approaches with processing
of SCADA databases and with modelling of economic loss with energy
not-supply.
As discussed in [MMH03], old “hardwired” solutions of the power systems
monitoring have many problems caused by their limited flexibility
which we are going to overcome by using the open and extensible multiagent
system implementation. This approach will be strong enough to
satisfy all our requirements.
5.1.1.1 The Multi-agent Approach
Rice is based on the principles of the multi-agent approach (see the Figure
5.1). Each facility of the power system is supposed to be autonomous
167
5.1.1 5. APPLICATION IN DIALOGUES – EPS SIMULATION
Figure 5.1: Role of supplementary agents in the Rice system
(self-deciding and autonomously reacting to system events) and so represented
by one software agent.
A general description of an agent-based software engineering methods,
that form the basis of the presented system, can be found e.g. in [Jen00].
Case studies of actual applications with agent-based architecture in the
industrial control systems are presented in [JB03].
Key idea of the multi-agent system is the idea of the agent itself.
There is plenty of definitions what agent is or should be, we can emphasize
some common features of agents:
• agent is cognitive. It watches its surroundings and makes its own
“cognitive map” of the world. All events taking place in the agent’s
environment are filtered by the agent according to their importance
for agent’s aim and important ones are stored in the agent’s knowledge
base (memory).
• agent is autonomous. There is no prepared plan of actions agent
should do to achieve his aim. In every time slice agent itself chooses
the best plan according to his knowledge base.
• agent is communicative. It can ask any other agent for any information
he needs to fill his knowledge base. Communication patterns
(i.e. topology of the communication net) are strictly dynamic. The
168
5. APPLICATION IN DIALOGUES – EPS SIMULATION 5.1.1
structure is built up ad-hoc according to agents’ needs for some
knowledge. There are no “hardwired” communication paths.
• agent is social. Main features of multi-agent systems are functions
of the system as a whole, not any particular agent.
In the Rice system, every power system facility is represented by one software
agent. Agents interaction standard is defined by the set of messages
agent must understand. These message sets are hierarchically arranged
so they grow up into a tree of agent types.
Every agent is therefore viewed as a black box with specified inputs
and outputs. To fulfill the idea of multi-agent system, every agent decides
on his state only from the state of his inputs. This leads to the
local-interactions-only system. And further: no information passing between
agents is obligatory. Every information must be obtained by the
question passed to the particular agent (one question can lead to a timebased
or event-based series of answers). This rule asserts that there is no
superfluous flow of information.
The “1:1” design (one facility = one agent) of the multi-agent system
is necessary for interaction with SCADA databases, as we will see further.
It is not easy to track events in dynamically changing environment.
Because we are in the situation where we do not know precisely what
variables are of interest for exploring the facilities failure rates (and thus
also for maintenance necessity and cost), things are even more difficult.
One possible approach is to experiment with the auto-adaptive networks.
The topology and information flow within the system is not fixed.
The leading wire agent can either copy the incoming energy flow from its
input to the output or it can react to demands from the respective power
network parts. The network is also dynamic as to the topology. The
topology is built up by series of question and answer couples. When a
connection between two agents is no more needed (e.g. because the power
line connecting them is down), communication is shut down and if there
is another agent which can facilitate the same function (and so restore the
energy distribution), new connection with it is established. The result
is that the system can be used for simulating ever-changing emergent
environment and for measuring the efficiency of the energy supply in
such an environment.
The application environment is very heterogeneous. It is absolutely
necessary to use technologies independent on hardware platform and op-
169
5.1.1 5. APPLICATION IN DIALOGUES – EPS SIMULATION
a)
def rLAgentStatus(self,status):
if self.msg.inReplyTo=="inputStatus":
print ’Input status is: "\%s"’\%status
if (status=="error"):
print "My input is in error state - switching to UPS"
b)
def rLAgentStatus(self,status):
if self.msg.inReplyTo=="inputStatus":
print ’Input status is: "\%s"’\%status
self.setRLStatus(status)
Figure 5.2: Reactions to zero voltage on the input line
erating system. The framework also supports wide range of programming
languages (due to CORBA). In the contemporary Rice prototype Python
and Java are used.
The main principle of the Rice system is that each agent can be implemented
by a stand-alone process or even stand-alone computer in a
computer network. If the communication flow between particular agents
is too large, it is possible to implement two or more agents within one
process and thus lower the communication overhead.
All agents in the system have the capability to communicate through
the CORBA and KQML protocols (see the Section 5.1.1.2). Each agent
has a defined type (denoting a set of messages which are accepted and
understood by the agent) and it is identified within the network by its
distinct name and a given identification number. The agent types are implemented
as a hierarchy in which each level is assigned a set of mandatory
KQML messages which every agent of this type must understand
and must be able to respond to it. The hierarchy looks like that:
• Agent
– Real-life agent (represents a particular power system facility)
∗ Line
∗ Facility
· Source
170
5. APPLICATION IN DIALOGUES – EPS SIMULATION 5.1.1
· Transformation station
· Switching station
· Consumer
– Organizational and auxiliary agents
∗ Register
∗ Helper
∗ Viewer
Each agent that belongs to some category must be able to respond to
every message from the defined set of messages. For example, agent of
the type Source must be able to respond to all messages mandatory for
types Source, Facility, Real-life agent and Agent.
Register The Register agent holds a database of all agents in the system,
their name, type and Agent ID (AID). Every agent must notify the
register about its existence before joining the network. The register then
passes a communication key to the agent. This approval for the network
joining serves as the certificate of authenticity and can be validated by
each agent in the network. The key also plays a role of an encryption
key for securing inter-agent communication. Agent is also given an AID.
The register agent can return an agent network address as the response
to a query based on agent AID, name or type.
Helper The Helper agent implements all global functions that are not
bound to any particular agent but are relevant to the system as a whole –
e.g. system identification values, supported protocol versions, description
of the allowed agent types, global configuration constants etc.
Viewer The Viewer agent performs a graphical display of the events
that take place in the system (see the Figure 5.3). There can be plenty of
viewer agents and each of them can visualize the state of the system in a
different manner – e.g. to plot a graph or to save the data in a database.
Viewer agent connection to the network is performed strictly dynamically
– after joining, the Viewer asks the Register for notifying about network
addresses of all viewable agents (KQML performative subscribe)
and then asks each of the agents to constantly provide information needed
for visualization of the agent (the same performative).
171
5.1.1 5. APPLICATION IN DIALOGUES – EPS SIMULATION
Figure 5.3: Example screen of the Viewer-type agent
Real-life agents In the sense of multi-agent systems theory, each part
of the system is an agent. By the term “real-life agent” (RLA), we
mean an agent representing a particular power system facility (it has an
equivalent in the real life). RLAs have (as opposed to other agents) a
huge set of mandatory messages regarding the power flow description.
Line A power line is the only part of the network which builds up
the network topology (i.e. its interconnection). Each line is labeled with
names of the two agents representing its input and output. Other RLAs
(of the facility type) are passive – they wait for the connection and
do not initialize it in any manner. So the topology of the RLAs remains
stable in the whole run which corresponds to the reality of the power systems.
On the other hand, the connection with other agents is dynamic – it
172
5. APPLICATION IN DIALOGUES – EPS SIMULATION 5.1.1
Figure 5.4: Communication layers
is possible to assign new monitoring agents to a running system.
5.1.1.2 Communication between Agents
The basic assumption of the system topology is that agents can be located
on separate machines or at least separate processes within one machine.
Hence the communication between them must be constituted by some
kind of network protocol – in our case it is TCP/IP. But this protocol
represents only the lowest layer of the communication which is extended
by three other protocols on the top two levels of the OSI model [OSI84] –
they are CORBA, KQML and the content language itself.
The first layer, TCP/IP is the standard Internet communication protocol
allowing the system to consist of geographically dissipated components.
The idea is that some of the simulating agents could be in future
replaced by on-line sensors monitoring a real facility operation. This
change is seamless and does not demand deep code intervention because
of the multi-agent architecture and the KQML flexibility.
The second layer, CORBA (Common Object Request Broker Architecture
[COR04]) is a standard architecture for inter-process (and even
inter-machine) procedure calling, used widely in system integration solutions
and decentralized software. It is strictly object-oriented. The
CORBA object programming interface is defined in an abstract Interface
Definition Language (IDL [IDL02]) which can be automatically translated
into many real programming languages (currently, we use mappings into
Java [CJa02] and Python [CPy02]). This language is formally similar
to structures used in data modelling meta-language UML (Unified Mod-
173
5.1.1 5. APPLICATION IN DIALOGUES – EPS SIMULATION
elling Language). IDL also includes standard specifications of general
data-structure mappings into many programming languages. Thus it is
not difficult to integrate code pieces written in different languages and
running on different operating systems and machines. All CORBA operations
all strictly platform independent – in the sense of hardware and
software equipment of the machines.
A big advantage of CORBA is the fact that objects communicate
with each other in the same way as when they are in the same process
– CORBA compliant ORB (Object Request Broker) facilitates all
data transformations and network transportation absolutely transparently.
This feature is especially important for possible system rearrangements
for the performance enhancement.
Another very important feature of CORBA which is used in our framework
is so called CORBA Naming Service (NS). This service is similar
to the Internet Domain Name Service, DNS. CORBA NS converts an
object name to a distinguished network address. Thus the knowledge of
an agent name and an address of the CORBA NS server is sufficient for
establishing a connection with the agent. NS server provides an agent
address on demand. This way it is possible to form an architecture which
can be subject to changes and is absolutely independent on the part of
network the agent is physically running. This key feature makes it possible
to reach the required system flexibility. Agents need to know only
the names of agents they are about to communicate with – all other
information is gathered dynamically.
The third layer, KQML (i.e. the Knowledge Query and Manipulation
Language [FFMM94]) is a language developed specially for usage in
multi-agent systems. It is an abstract definition of the message purpose,
not the message content itself. This way, the meaning of the message is
transparently separable from the message object, which is very useful for
system integration and definition of abstract message types. The detailed
message flow schema between two agents is depicted in the Figure 5.5.
The last layer is the message content itself. Thanks to the KQML
structure, it can be defined in many languages with arbitrary structure.
The message content language used can even vary between individual
messages and can be dynamically negotiated by the communicating
agents.
For instance, a definition of a data structure and an interface of an
agent which is able to communicate in KQML looks in IDL this way:
174
5. APPLICATION IN DIALOGUES – EPS SIMULATION 5.1.1
Figure 5.5: The Implementation of the Rice Communication Schema
struct KqmlMessageStruct {
string performative;
long sender;
long receiver;
/* ... */
any contents;
long inReplyTo;
string force;
};
interface KqmlReceiverInterface {
exception badKQMLMessage{};
// send KQML message - standalone args
void kqmlMessageArgs(in string performat, in long sender,
in long receiver, in any contents, in long inReplyTo,
in long force)
raises (badKQMLMessage);
// send KQML message - struct
void kqmlMessage(in KqmlMessageStruct message)
raises (badKQMLMessage);
};
175
5.1.1 5. APPLICATION IN DIALOGUES – EPS SIMULATION
For each of the supported languages, there is an IDL compiler which
builds a skeleton in the desired language. The skeleton corresponds to
the structures defined in IDL (in the above example, it will build the
KqmlReceiverInterface in the Java language).
The KQML layer is built above the CORBA layer which facilitates
the (virtualized) connection between agents. The KQML language is
based on the linguistic theory of the Speech acts [Smi90], published by
Searle in [Sea69]. One of the basic ideas of this theory is that every
communication act can be categorized as an announcement, a query, a
demand etc.
The KQML communication is strictly divided into two levels – level
of the speech act resolution and the message content resolution. For each
speech act type there is one or more so called “KQML performatives.”
Performatives provide basic information about what type of information
or action an agent demands.
The KQML language is defined as an abstract query and manipulation
language. The basic concept is that each of the communicating agents
has its own knowledge database concentrating knowledge about the outer
environment and the agent itself. But this is only a formal concept. In
reality the agent can work even strictly reflexively – but it must represent
its actions as manipulations with the knowledge database. For this reason
we call the desired data structure on which the KQML operates a Virtual
Knowledge Base (VKB).
KQML has two fundamental concepts: (1) each agent is autonomous
and only it decides what to do in a particular time, (2) agent A can ask
agent B for information from its VKB. KQML contains a set of performatives
which express a desire of an agent A for agent B to make an
effort to achieve some state of the environment.
The KQML message envelope defines from whom, to whom and when
was the message delivered and in which form the sender expects an answer.
The content itself is communicated in the content language, which
can be any possible language which is adequate for the message content
expressing and which is understood by both the agents. In practice,
usually Prolog or KIF1
are used.
For an illustration of the inter-agent communication we take a typical
situation in a power system simulator (the example is simplified for better
1Knowledge Interchange Format [GF92] was designed especially for this purpose.
176
5. APPLICATION IN DIALOGUES – EPS SIMULATION 5.1.1
Figure 5.6: Example of EPS agents’ communication
clearness): agent A represents a transformation station transforming a
very high voltage to a high voltage. Agent A is connected to agent B
which represents high voltage line and which is connected to agent C
representing a distribution point (see the Figure 5.6).
Agent B asks agent A for a notification of every A’s state change and
every A’s output voltage change:
(subscribe
:sender B
:receiver A
:timestamp 1113340454
:reply-with query_1
:language KQML
:ontology KQML_ontology
:content (ask
:sender B
:receiver A
:in-reply-to query_1
:reply-with 2
:language Prolog
:ontology Power_system
:content out_voltage(X), state(Y)
)
)
As soon as the agent forms a KQML message, it contacts the CORBA
Naming Service and asks it for a network address of the recipient (in the
177
5.1.1 5. APPLICATION IN DIALOGUES – EPS SIMULATION
case that it is not in his VKB yet) and contacts the recipient (again,
in the case that it has not done this already before). The sending of
the message itself is realized by a CORBA call which is in the formal
sense identical to a call of a function in the same process (programming
language “native” call).
By this technique a dynamic connection between the agents A and
B is established (i.e. the output of the device represented by the agent
A is connected to the input of the device represented by the agent B).
Similar connection is established between the agents B and C.
If there is an outage on line B and the current supply is canceled,
B sends to C a message of this format:
(tell
:sender B
:receiver C
:timestamp 1113341454
:in-reply-to query_2
:reply-with query_3
:language Prolog
:ontology Power_system
:content out_voltage(0), state(fatal_failure)
)
C reacts to this message in this way: it sets the voltage of all its outputs
to 0 V too or switches to another (backup) input line.
The big advantage of KQML is that the agent B can send a message
with the same format to the visualization agent or to an agent which loads
data into the failure database. It is not necessary to develop a specialized
protocol for each communication channel.
It is also possible to change the parameters of the communication
with only a little change of the message content. For instance, in the
process of building a comprehensive database of power supply failures
by gathering information from the agents, the central database requires
more detailed data – so the corresponding agent’s subscribe message
will contain other queries – the content field will in this case look like
this:
:content state(fatal_failure), facility(U), type(V),
serial_num(W), owner(X), locality(Y),
178
5. APPLICATION IN DIALOGUES – EPS SIMULATION 5.2
description(Z)
The type of the speech act has not changed so even the KQML envelope
remains the same. Only the message content is different.
An important question, of course, is whether the described four-level
communication framework will have enough bandwidth for the power system
simulation. A quality agent’s design should be based on the following
principle: “Agents are communicating only those data which they really
need” – i.e. as much data as possible is processed locally by the agent.
The other agents should ask only for results of this processing. E.g. the
running temperature of a facility can significantly affect the reliability of
the facility but no agent should ask for the temperature to deduce the
facility’s state – instead it should ask for the conditions of the facility or
the probability of a failure in the following time period. The amount of
the communicated data can be significantly reduced in this way.
The content language should be a language which is a “native” language
of an agent. It is very inefficient to choose a language which demands
a conversion of existing data to another format and back just for
the sake of the communication.2
5.2 Human-machine Dialogues with the Rice
System
For thousands of years natural language (NL) dialogue has been the
most traditional way of communication with humans. Therefore, following
this tradition, NL dialogue forms also a new and efficient way of
communication between human and computer. NL interfaces provide successful
systems for help systems, secretariat systems or tool for complex
systems analysis. The purpose of these systems is to handle restricted
and well-known domains. In these domains, the designers can express the
rules of understanding user’s requests and generating system’s responses.
A real electrical power network represents a complicated system of
large amount of cooperating appliances. The analysis and prediction of
failures in such systems (see [KMG01, PSH01]) is a difficult task that
often suffers with overspecification problems. The aim of our research
2here we speak of communication between agents, not human-machine communi-
cation.
179
5.2.1 5. APPLICATION IN DIALOGUES – EPS SIMULATION
is to explore the possibility of enhancing the communication with such
complex systems analysis by means of natural language dialogue inter-
face.
Methods of automatic analysis of domain-specific texts heavily exploit
language resources which are adapted to the respective domain. The creation
of the first phase of a basic resource for analysis of dialogues about
electrical power networks was published in [PH05a]. The size of the textual
corpus described there was about 100 000 token, the morphological
tagging of the corpus was ambiguous and the coverage of the syntactic
analysis on the corpus was about 50 %. In this text, we show the progress
in creating a corpus which is 10 times bigger, does not contain ambiguous
morphological tags and the quality of the relevant syntactic analysis has
increased up to 76 %.
5.2.1 Building a Specialized Corpus
In this section, we describe the procedure of collecting the domain-specific
texts.
As a starting point, we used a set of 14 documents from the application
domain – reports, program documentation and technical specifications of
electrical devices. These texts together contain approximately 100,000
positions [PH05a]. These texts served for providing a basis for the WebBootCat
[BKPR06] tool3
for extracting keywords and searching the web
pages to find and download similar documents.
As a result, we obtained a corpus of about 1 million tokens (words and
punctuation). The newly extracted key terms from this corpus contained
terms from the electrical power networks domain (e.g. “elektrické zařízení”
– electrical device, “elektrická energie” – electrical energy, “napájení”
– supply, “proud” – current) as well as more general words such
as “material”, “producer” or “system” (see the Table 5.1). The key
terms search worked also as a check that the system downloaded just the
domain-specific documents. A detailed description of the corpus can be
seen in the Table 5.2.
3developed also in the MU NLP Centre
180
5. APPLICATION IN DIALOGUES – EPS SIMULATION 5.2.1
Table 5.1: Most frequent keywords extracted from the final corpus
word frequency
zařízení (device) 5 983
musí (must) 3 797
napětí (voltage) 2 919
proud (current) 1 945
elektrické (electrical) 1 859
energie (energie) 1 813
soustava (system) 1 584
měření (measure) 1 484
vedení (wiring) 1 481
části (parts) 1 011
těchto (these) 1 027
provoz (operation) 974
použití (usage) 809
zdroj (source) 759
Table 5.2: Characteristics of the electrical power networks domain corpus
Number of tokens 1 034 511
Number of sentences 43 738
Number of documents 160
Median number of words in a sentence 15
Average number of words in a sentence 24
181
5.2.2 5. APPLICATION IN DIALOGUES – EPS SIMULATION
5.2.2 Morphological Tagging
To be able to analyse the corpus sentences with advanced language processing
tools, we need to supplement the corpus by lemmatization and
morphological tagging. For this task, we have used the Czech morphemic
analyzer ajka [Sed05] that provides lemmas and all possible morphological
tags for more than 6 million Czech word forms. For the tag disambiguation,
we have used a statistical disambiguation tool [Š04] that is
being developed at the MU NLP Centre. These two tools together form
a Czech morphological tagger.
The overall morphological tagging coverage on the electrical power
networks domain corpus was 95.84 %. The most frequent expressions that
were not recognized are the domain-specific words (e.g. “rezistance” –
resistance), abbreviations (e.g. “VN” – stands for “vysoké napětí”, high
voltage) and physical units (“kVA”, or “kPa”). For successful processing
of natural language texts the mentioned morphological analysis tools need
to be supplemented with these terms.
5.2.3 Syntactic Analysis of the Domain Texts
Since the syntax analysis is one of the basic steps in natural language
understanding, we plan to use the synt parser4
as a part of the developed
dialogue interface. Therefore, it is important to know, how successful is
the parser on the texts from the particular domain and what specific
phenomena need to be added to the system for good-class analysis of
these texts.
In the following experiment we investigated the parser coverage in
case of domain-specific texts.
Within the experiment, the whole corpus was split into sentences, using
the default sentence boundaries provided by the WebBootCat system.
Then, the synt system was executed on these sentences and the number
of accepted sentences was computed. As can be seen in the Table 5.2,
the total number of sentences in the corpus is nearly 44 000. The results
of the experiment can be seen in the Table 5.3. In the Figure 5.7, there
is an example of a synt derivation tree for one of the corpus sentences.
The Table 5.3 shows that the parser has accepted about 62 % of the
corpus sentences, which is much less than in the case of common sentences
4see the Chapter 3
182
5. APPLICATION IN DIALOGUES – EPS SIMULATION 5.2.3
Table 5.3: Results of syntactic analysis with the synt system
Sentences accepted 27 209
Sentences rejected 16 529
Parser coverage 62.2 %
Average parsing time 1.96 s
start
sentence
clause
np
np
np
N
´Upravu
np
np
ADJ
jalov´e
np
N
sloˇzky
np
N
proudu
pp
PREP
pomoc´ı
np
np
N
proudu
np
N
kondenz´atoru
V
naz´yv´ame
np
ADJ
paraleln´ı
np
N
kompenzac´ı
ends
’.’
.
Figure 5.7: Example of a derivation tree for one of the corpus sentences.
183
5.2.4 5. APPLICATION IN DIALOGUES – EPS SIMULATION
occurring in a reference corpus. It can also be seen that the average
parsing time is significantly higher than the parsing time measured on
the reference corpus. One of the reasons for this fact is the high average
sentence length (see the Table 5.2).
The main reasons for the low coverage of syntactic analysis of the
corpus are as follows:
• the frequency of yet unknown key terms is higher in the domainspecific
texts
• special expressive techniques of the technical texts (frequent use of
abbreviations, table and figure references, numbers, equations, . . . )
• possible errors in the sentence-splitting mechanism in the WebBootCat
system
An example of difficulties described in the first two points can be illustrated
e.g. with the sentence: “Celkový instalovaný příkon cca 6 x 400 =
2400 kW.” (Total energy input installed (is) cca 6 x 400 = 2400 kW).
In order to verify the last of the previous three points, we have made
an additional experiment. We found out that the corpus contains (among
others) many very long sentences (up to 1 000 words), so it is evident that
some sentences are marked badly. Therefore, we filtered out the sentences
with a high number of words and measured the parser coverage again.
The results are shown in the Table 5.4. At shorter sentences (that are
more probable to be marked correctly), the parser coverage is significantly
higher, up to 76.4 %. These numbers show that some more fine-grained
sentence splitting procedures (than the default one that is integrated in
the WebBootCat system) should be applied to this kind of corpus.
5.2.4 Designing an Intelligent Dialogue Interface
Dialogue interfaces are interfaces that use natural language for communication
with users. Designing an intelligent dialogue interface is a complex
task that covers several fields in the area of human language technology,
including morphological, syntactic and semantic analysis, information retrieval,
discourse analysis and language generation. Designing such complex
system with full functionality for general use is unfeasible with the
current level of technologies. Nowadays, such systems are always domain
dependent, i.e. tailored for a particular domain.
184
5. APPLICATION IN DIALOGUES – EPS SIMULATION 5.2.4
Table 5.4: Parser coverage measured on the sentences with limited length
sentence length number of parser coverage
limit (words) sentences (percent)
None (all sentences) 43 738 62.2
40 36 043 68.5
30 32 219 72.0
20 24 430 76.4
At the first stage, we have focused on development of a question answering
system for searching in large databases of complex facts called
UIO. Currently the system is able to process natural language queries
over the administration system of the Masaryk University – it contains
data about 35 000 students, their studies, courses and other information.
We are now adjusting the system for queries over the database of electrical
network failures. The UIO system contains a general algorithm for
question analysis, information retrieval and answer generation and it is
being equipped with knowledge about the domain of electrical network
failures, nowadays.
5.2.4.1 The Question-Answering System UIO
UIO is a multi-domain oriented question-answering system for the Czech
language incorporating advanced natural language processing techniques,
designed and developed by Lukáš Svoboda in the MU NLP Centre. A simplified
scheme of the system is depicted in the Figure 5.8.
The analysis of Czech sentences has to cope with practically free word
order and a high degree of ambiguity caused by the inflectional nature of
the language. UIO tries to solve these problems by using partial analysis
with known question schemata.5
First, the morphological analysis of
the user question is performed, which finds the corresponding lemma(s)
and part-of-speech tag(s) of each word. The morphological disambiguation
is usually not performed, the question is analysed with all possible
5Please note that the description is simplified, detailed information can be found
in [SP04]
185
5.2.4 5. APPLICATION IN DIALOGUES – EPS SIMULATION
Figure 5.8: UIO flow chart.
lemma/tag combinations. Specific part of this phase of the analysis lies in
recognition of multi-word expressions (MWE). For UIO, a special MWE
module has been created [Svo04]. This module works with a database of
MWEs (containing more than 100 000 items) and provides functions for
MWE inflected forms and lemmatization.
The result of the morphological analysis is used as the input for the
syntactico-semantic analysis of the question. For this purpose, a special
semantic-based grammar is used. It is based on a context-free grammar
which is augmented with tokens carrying information about the semantics
of the matched grammar. In addition to this, the grammar is extended
with special features considering the needs of inflectional languages:
a) besides the usual exact match of terminals, matching based on equal
lemmas is supported in order to be able to face the problem of
various inflected forms.
b) in order to face the free word order problem, matching any permutation
of the given tokens is supported.
c) to improve the flexibility, matching of synonyms is supported.
186
5. APPLICATION IN DIALOGUES – EPS SIMULATION 5.2.4
{
domain => ’ElectricalNetworksFail’,
from_date => { month => 1,
day => 1,
year => 2007 },
to_date => { month => 1
day => 31,
year => 2007 },
short_circuit => true,
type => ’ListFail’,
}
Figure 5.9: Semantic representation of the utterance “Vypiš všechny
vzniklé závady v lednu 2007, kde nastal zkrat.” (Show all breakdowns
in January 2007 where short circuit happened.)
UIO uses a bottom-up chart parser that is able to work with the grammar
extensions mentioned above. The result of the analysis is a set of
parse trees corresponding to various possible ways of parsing the question.
A probability of correctness of the analysis is associated with each
tree. The semantic representation of the question is constructed from the
tokens carrying the semantic information. The semantics is represented
as a list of semantic attribute/value pairs with possible hierarchical organization.
Several special semantic pairs are available for processing of
questions containing superlatives (What is the highest...). An example
of the semantic representation is depicted in the Figure 5.9
The bottom-up chart parser used in UIO is also able to process partial
parsing which is of crucial importance since it happens very often that
the question does not match the grammar and cannot be fully parsed.
In such case, the parser returns a set of parse trees corresponding to
the successfully analyzed parts of the question. From the parse trees,
partial semantics of the question can be computed. The capabilities of
the system can be presented by examples of different questions that the
system is currently able to process and answer – see the Figure 5.10.
Based on the semantics of the question and the associated probability
of correctness, UIO first determines the domain corresponding to the
187
5.3 5. APPLICATION IN DIALOGUES – EPS SIMULATION
question. In the case that the question matches the grammar, this problem
is trivial since the domain is associated directly with the root rule
of the matched phrase. If only partial semantics is available, the most
probable domain is inferred from the available information. Sometimes,
there are multiple domains corresponding to the same phrase (e.g. a time
table inquiry might refer to a bus connection as well as a flight). In such
situations, the answer is extracted for all the domains.
As soon as the domain is determined, the semantic representation of
the question is passed to an answer extraction algorithm. The algorithm
either detects that attributes needed for answering the question are missing,
or extracts the answer from the information source predefined for
each domain. The information source can be a local database or a web
server that provides the required information. The extracted answer is
converted into a semantic representation similar to the one used for representing
the question. This representation can be further processed – if
multiple domains match the question and there is an empty answer for
some of them, it can be ruled out, or, in the case of questions containing
superlatives, the answer must be computed from the semantic representation.
Then, the result is passed to a presentation module that transforms
the information into a text or a table that can be presented to the user.
5.3 Rice Usage Scenarios
There are many ways of possible usage of the Rice system but all of
them can be summarized to four main usage patterns. Typical usage
of the Rice system will consist of several simulation runs with different
properties of the simulated environment and examining the impacts of
the changes. The Rice usage scenarios can be grouped according to the
examined subject and the means of the environment manipulation:
Global Manipulation, Local Effect The most common usage of the
system will consist in the comparison of the behaviour and characteristics
of two or more power systems consisting of the same basic
components but different interconnection between them (analysis
of different power system topology impacts).
In the Section 5.4, the results of an experiment with simulation
of the economic loss caused by energy not supply are presented.
188
5. APPLICATION IN DIALOGUES – EPS SIMULATION 5.3
Jaké byly příčiny nahodilých poruch v roce 2003? (What were the causes of
incidental failures in 2003?)
Kdy bylo vyrobeno zařízení, u kterého se vyskytla nahodilá událost 15.1.2003?
(When was the device, which was the source of an incidental event on January 15,
2003, produced?)
Jaká je nejčastější příčina události v roce 2003? (What is the most frequent
cause of events in 2003?)
Vypiš poruchy, které se staly 16.1.2003 na vodičích (List failures that happened
on January 1, 2003 on leading wires.)
Jaké druhy zkratu nastaly v roce 2003 na stožáru? (What kinds of short circuit
appeared in 2003 on a pole?)
Jaký druh sítě byl v roce 2002 poškozen nejvíce? (What kind of network was
the most damaged in 2002?)
Kdy mělo poruchu zařízení vyrobené společností ESB Brno? (When did a device
produced by ESB Brno company fail?)
Jaké je nejporuchovější zařízení vyrobené společností Kablo Kladno? (Which
of the devices made by Kablo Kladno has the highest failure rate?)
Kolik distribučních transformačních stanic bylo mimo provoz 16.6.2001? (How
many distribution transformation stations were out-of-order on June 6, 2001?)
Jaká zařízení s napětím 400V byla postižena nahodilou událostí? (Which devices
with 400V voltage were affected by an incidental event?)
Jaké se staly události na transformátoru napětí? (What events happened on a
voltage transformer?)
Vypiš data, druh události a zařízení se záznamy z 16.1.2003. (Show the dates,
event types and devices with records from January 16, 2003.)
Kdy a na jakém zařízení došlo k nahodilým událostem způsobeným zásahy
cizích osob v roce 2005? (When and on what device did an incidental event
caused by foreign persons happen in 2005?)
Jaká je průměrná délka události v roce 2003? (What is an average length of an
event in 2003?)
Jaké jsou příčiny událostí, které vedly v roce 2003 k poškození reaktoru?
(What are the causes of events leading to a reactor damage in 2003?)
Figure 5.10: Examples of NL questions that can be handled by the current
implementation.
189
5.3 5. APPLICATION IN DIALOGUES – EPS SIMULATION
Two power systems with slightly different topologies are described
there and the corresponding economic impacts of the outages on
particular network nodes are compared.
Such situation represents a typical example of the exploration of the
local effects of the global system change. We should note here that
the topology in Rice is also defined “from bottom up.” Every agent
has a definition of its neighbours and there is no overall topology
definition so any changes of the topology are easy to represent. No
intervention into the agent behaviour definitions are needed.
Local Manipulation, Global Effect The second usage pattern is similar
to the first one but one more Rice feature has to be used – the
data gathering and analysis tools (in the case that the explored
global effect is composed of individual local effects).
Rice includes a data gathering mechanism similar to the failure
databases used in the real power system monitoring. Each agent has
its own data storage called event database. All events relevant for
the particular exploration are stored in this database in a standard
form. The system includes a tool for merging these databases for
preparation of overall statistics such as the total outage time, the
outage-caused losses or the overall energy consumption.
An example of such an exploration can be also found in the Section
5.4.
Local Manipulation, Local Effect The third usage pattern is used
for exploration of the impact of a change in the behaviour of some
particular network component on other components. This action
cannot be reduced to a relation between two components because
the results of the change are often obtained by a non-trivial function
over a whole inner-standing network segment.
Tests of this kind are possible in Rice, however, no such test has
been realized yet.
Visualization As we have said already, Rice includes tools for exporting
events taking place in the network into a database. But this events
can also be imported on-line from the databases. With this feature,
Rice can be used for visualization of the processes taking place in
the real power systems.
190
5. APPLICATION IN DIALOGUES – EPS SIMULATION 5.3
Failure databases from real networks can be imported into Rice and
“replayed” in the simulated environment either as a visualization of
the EPN processes, or more complicated exploration of the failures
context (e.g. simulation of several ways of a failure recovery and
their effects).
The Rice system includes a rich library for support of fast system customization
and implementation of a particular power system. Parts of
this library implement basic power flow operations, standard agents behaviour,
etc. All inter-agent communications are based on the KQML
language.
In this section, we will shortly describe the basic principles of a power
system components implementation in the Rice system. We will use the
terminology common in the field of a multi-agent software development.
Almost every agent property is defined in the way how it changes
the agent’s interaction and communication (“behaviour”) with its surroundings.
A typical agent definition consists of three basic classes of its
behaviour:
Event-driven actions are triggered in the case of an important environment
change. For our topic these kind of events are important:
energy flow (energy demand) change, facility error (resulting in outage
or not), outage caused by other facility error. To satisfy the
multi-agent design principles and make the development process
clear, we must distinguish well before the two latter events.
Note that there is no mechanism to simulate the outage propagation
as is. This is not needed and should not be implemented. Instead,
we can see every agent as an autonomous entity which reacts to
the event “my input lines are at zero level.” The simplest reaction
is to propagate the outage state to all agent’s outputs, but more
sophisticated agents that will search through their inputs and ask
every peer if it can supply more power to overcome the outage can
be implemented.
Using complicated behaviour implementation a very complex and
even conventionally hardly computable error propagation throughout
the power system can be modelled even if the behaviour of all
engaged agents can be highly dynamic and state-dependent. For
191
5.3.1 5. APPLICATION IN DIALOGUES – EPS SIMULATION
Figure 5.11: Simple bus
instance the topology of the network can vary in the dependence of
the power lines load, daytime, etc.
Time-based actions are special types of the event-driven actions. They
are triggered (1) by approaching an absolute time or (2) after expiration
of a time period (relative time). In the real implementation
the model must incorporate various time-based actions: “Service
period reached,” “Critical outage time reached” or “End of the fiscal
period reached.” All of them (plus some others) have built-in
support in Rice.
Idle-loop actions (ILA) are actions which should be performed continuously
(or as frequently as possible). Note that Rice does not
have discrete time (time flow in exact steps) and so time between
ILA triggers can vary. If the simulation needs exact time slices, the
time-based actions with simulator time speed appropriate to the
hardware performance are to be used.
5.3.1 Example Nodes Implementation
In this section, we will demonstrate the process of an agent behaviour definition
on two examples. We will concentrate on the process philosophy
and marginally on the practical implementation and particular code.
5.3.1.1 Simple Bus
In the first example, we will demonstrate the main guidelines of a simple
bus agent definition (the Figure 5.11). The bus has one input and three
outputs.
192
5. APPLICATION IN DIALOGUES – EPS SIMULATION 5.3.1
Power demands originating from outputs are sent to the input. In
case that the demand exceeds the capacity of the source, the demanding
agent must be informed that the demand can not be satisfied. The agent
will not simulate over-voltage.
Outage on the input will be propagated to all outputs. There is no
backup so no special action will be taken.
We will define two time constants as a maintenance period and maintenance
duration. Before the maintenance the agent will send a message
to its outputs that there is a maintenance outage scheduled. This makes
it possible for the outputs to establish any other power source, intelligently
change the network topology and overcome the outage.
According to this design plan we must define these event-driven ac-
tions:
1. power demand arrived from output: check if the saturated demand
will exceed the overall demand limit. Send granting or denying
message.
2. outage on input: send outage message to all of the outputs and set
the agent to the “error” state.6
The first action will be defined by only a few lines of code.7
Rice has a
built-in KQML parser which will launch the agent function corresponding
to the arrived message. Function powerDemand is launched when the
power demand KQML message arrives. The code follows:
def powerDemand(self,msg,demand):
if self.powerConsumed + demand > self.maxDemand:
utils.kqml.send(msg.sender,
PyMessage("receiver.denied()"))
else
self.powerConsumed+=demand
utils.kqml.send(msg.sender,
PyMessage("receiver.granted()"))
6There is a principle in the multi-agent programming saying that any agent should
not set the state of another agent. Instead of this every agent decides on its own about
its state. We silently conform to this notion here: sending the outage message to the
outputs does not mean that the output agents will be set in outage state too. They
can have some mechanisms to overcome the outage (power backup, intelligent source
searching etc.) and so they decide how to react to this message on their own.
7in the Python programming language
193
5.3.1 5. APPLICATION IN DIALOGUES – EPS SIMULATION
All KQML message constitution, sending, delivering and parsing is managed
in the Rice standard library. All we have to define is just the core
logic.
The second action will be defined in this way:
def rLAgentStatus(self,status):
if self.msg.inReplyTo=="inputStatus":
print "Input status is: ’%s’" % status
# no power backup source here – just copy the input status
self.setStatus(status)
# inform all outputs about status change
for sub in self.outputSubscriptions:
sub.reply("receiver.rLAgentStatus(’%s’)" % status)
rLAgentStatus is a function launched when the message about the agent
status has arrived. By examining the in-reply-to KQML field we verify
that the status change is related to the input (in this example it is not
really necessary but it is in the cases where the agent has more inputs).
For informing the outputs we use standard KQML tool – subscriptions.
The agent that wants an information about status changes from
some other agents sends a subscription demand for status message to that
agent. Subscription message means that the agent wants to be informed
about all changes of some value in the other agent (in this case the status).
So when a status change occurs, the agent informs all agents which
subscribed for this kind of information. This mechanism realizes other
principle of the multi-agent design: only the demanded (and so really
needed) information are communicated between agents.
The last function in the example is the time-based maintenance. The
implementation uses Rice built-in functions.
def maintenanceUpcomming(self,status):
for sub in self.outputSubscriptions:
sub.reply("receiver.rLAgentStatus(’anticipating outage’)")
# wait for a defined time before the actual maintenance procedure
utils.timer(self.maintWaitTime,self.maintenance)
def maintenance(self):
for sub in self.outputSubscriptions:
sub.reply("receiver.rLAgentStatus(’outage’)")
The function maintenanceUpcomming is launched in regular time periods.
It informs all outputs about the scheduled maintenance and sets the
194
5. APPLICATION IN DIALOGUES – EPS SIMULATION 5.3.1
Figure 5.12: Duplicate busbar
timer which will launch the function maintenance after the chosen time
during which the agents connected to the output can look for a backup
power source.
5.3.1.2 Duplicate Busbar
In this example, we have a typical duplicate busbar with two inputs and
three outputs, see the Figure 5.12. Besides all actions defined in the
previous example, we will provide a new action: if there is an outage
on one input line, all outputs are switched to the other line (maximal
demand checking is similar to the first example).
The principle of the agent implementation is similar to the previous
one. With a minor customization we can use all actions defined above.
The difference is mainly in the need to distinguish the messages originating
from the first and the second input. We will use the reply-with KQML
field for it. The subscription message will be sent by this message:
utils.kqml.send(self.input1,
PyMessage("receiver.rLAgentStatus()",
msgType=utils.kqml.subscription,
replyWith="input1Status"))
By this message the agent subscribes to the information about all status
changes of the agent input1. Incoming responses to this subscription will
be tagged with the string input1Status in the in-reply-to KQML field.
Reaction to the outage on one line – switching all outputs to the
second one will than be defined by this code:
195
5.4 5. APPLICATION IN DIALOGUES – EPS SIMULATION
def rLAgentStatus(self,status):
if self.msg.inReplyTo=="input1Status":
self.allToLine2()
if self.msg.inReplyTo=="input2Status":
self.allToLine1()
Function allToLineX will check if the overall demand does not exceed the
maximal allowed demand and eventually propagate the outage similarly
to the first example.
5.4 Modelling of Economic Aspects of a Power
System Failure
Each electrical power system (EPS) failure causes a chain reaction of
effects, all with negative economic impacts. The effects can be categorized
a) from the point of view of the supplier and b) from the point of view of
the customer. Both of these groups of effects can be further divided into
parts denoted as direct and indirect costs [BA96]. The overall amount
of costs arising as a result of a power system failure can be calculated as
sum of all these costs. However, specification of these costs is a long and
complicated process, since all these four groups consists of many items
such as the cost of technology repair, the cost of unsold energy, the cost
of technology restart (on the customer’s side) or the cost of lost profits.
The ability to provide a prediction of these costs for a particular power
system failure is thus a very welcomed feature of any computer system.
The Rice system is able to provide a flexible multi-agent dynamic network
platform for complicated simulation of all power network processes
including the setup and computation of possible costs of particular effects
of an outage within the network.
An interesting application of Rice comes out from the association
with classical SCADA databases. These databases are primarily used
to determine particular facility failure rates, various conditions of the
failures, defective facility series etc. Such information is processed by
means of the classical data mining and statistical analysis. In spite of
that data mining is relatively advanced field, it can be hard or impossible
to harvest some types of information from the database, where the entries
map only particular failures, outages, etc.
196
5. APPLICATION IN DIALOGUES – EPS SIMULATION 5.4
But SCADA can be used for “profiling” particular facility type and
its behaviour in the network which can be then used in the development
of an agent simulating this facility. The particular methodology of this
process must be defined yet and it is our plan for the near future.
We can also use a backward process. In Rice, agents send information
about their state to the application which visualizes processes taking
place in the network. The same principle can be used to create a simSCADA
– simulated SCADA. By comparing data in the real SCADA and
simSCADA, we can numerically evaluate the simulation accuracy.
Moreover, because failures are simulated by message passing, it is
relatively simple to track down the failure origin and join this original
failure with its consequences (measure precisely the amount of undelivered
energy). This could be hard with classical SCADA and we consider
this a promising feature of our system. It is simple to measure the overall
sum of the undelivered energy and use this measure for the optimization
of the agent behaviour principles.
The Rice framework offers several important features as regards the
modelling of economic aspects of failures in the power supply:
• reactive agents – agent behaviour is determined by its inner state
and its perceptions (i.e. the state of the agent’s environment).
• local interactions – interaction between agents is strictly in the
“peer to peer” manner. There is no global authority or global
synchronization.
• on-demand information – every interaction is initiated by an agent
that needs a piece of information for its operation. This has two
practical results: (1) At the time of the implementation we do not
need to know which information will be necessary for which agent
and so we do not have to pre-design data flows in the system. (2)
Data flow rate is on the least possible level.
• massive parallelism – every agent in Rice consists of one or more
threads or even separated applications. This means that no agent
can be blocked by misbehaved peer.
• non-discrete time – some multi-agent systems are designed as “stepbased”
applications. This means that agents must work with small
time slice operations and the whole operation flow is constructed
197
5.4 5. APPLICATION IN DIALOGUES – EPS SIMULATION
by a kind of time sharing algorithm. This causes serious practical
impacts on possible agents design, coding style and programmer’s
skill. As we have mentioned above, Rice is massively parallel and
so the agent design does not require to pay any special attention to
the time sharing.
• not necessarily real-time – Rice can be used as a real-time system
but a typical usage will be the simulation of a power system operation
in a long time period (months or years) within a short-time
period (accelerated simulation).
In this section, we will explain usage of the Rice system on the example
of modelling a hypothetical part of an EPS. The purpose of the example
is to present the process of a model implementation in the Rice system.
In the example we will need two kinds of actions: (1) computation of
actual outage cost and (2) computation of actual power demand.
The model has eight facilities: one power source (A), three transformation/distribution
points (B,C,D), four consumer points (E,F,G,H)
and seven power lines interconnecting the points (L1–L7). Visualisation
of the network in the Rice-Viewer is displayed in the Figure 5.13.
In addition to the network topology, we must define the following
behaviour principles for each agent:
• power demand function – amount of power consumed by the facility
in the time
• reliability function – frequency and duration of the facility errors
in the time
• reaction to inputs outages – means of reactions can vary. Basic
agents set their state to the input outage, set their output level
to zero power or start to malfunction. Intelligent agents can start
searching for new inputs configuration to re-achieve full energy supply
from other sources.
• loss function – defines the amount of immediate and cumulative lost
profits regarding the outages and other malfunctions of the facility.
As well as with other functions, the lost profits depend on the day
time, the day of the week, etc.
198
5. APPLICATION IN DIALOGUES – EPS SIMULATION 5.4.1
Figure 5.13: The example model of EPS used for computation of economic
losses.
5.4.1 Network Parts
The parts of our simulated power network are equipped with the following
parameters.
5.4.1.1 Source
In our example we consider a particular energy source A with constant
available energy (1.5 MW). Demands exceeding this limit will be simply
rejected (possibly causing outage in the demanding facility). The source
itself is maximally reliable (possibility of an self-caused outage is zero).
5.4.1.2 Distribution Points
We have three distribution points (DP) of different kind. First DP B
is considered to be duplicated and so inner-fault-free. Therefore its reliability
is maximal. Second DP C is dedicated to supply power for two
factories so its reliability must be at good level. Let it be one error a year,
with average duration of 20 minutes, Gaussian distribution of the outage
199
5.4.1 5. APPLICATION IN DIALOGUES – EPS SIMULATION
duration with standard deviation 5. But DP C has another limitation
– the total output power can be only 0.7 MW. Demands exceeding this
limit will lead to denying the demand.
Last DP D is an old local facility dedicated to city quarter F. The
D’s error level is high (4 errors a year). Time to recover is long (average
3 hours) and very varying (deviation 1.5).
5.4.1.3 Power Lines
In the Rice system, there are separate agents for each power line. In our
example we consider all power lines fault free. There is only one cycle
in the network graph: A-B-G-C-A, so there are two ways of delivering
power from A to G: A-B-G and A-C-G. The latter one is used as the
network initial configuration.
5.4.1.4 Consumer Points
There are four consumer points (CP). CP E and H are factories with
middle-cost production. Both factories use no machinery which can be
destroyed in the case of an outage. The lost profits of the factories can be
approximated with a linear function as loss of $3 000 per hour (factory E)
or $1 000 per hour (factory H). The average power consumption of the
factories is 100 kW for factory E and 200 kW for factory H. Power consumption
changes in steps with the Gaussian distribution with the mean
of 1 kW and the deviation of 0.1. Both factories do not run overnight, so
we must consider power consumption deviation throughout the day and
count a loss only when the factory is running (outages in the not-running
time are considered zero). Factory E has old internal power lines which
can cause rare internal outage (average rate is 0.1 per year, duration of
5 hours with the deviation of 0.01).
Factory G is larger than the previous ones and has outages-sensitive
high-cost technology. Even short outages can cause high money losses
because of damaged technology parts. That is why the loss function cannot
be approximated linearly. Instead, we will use the following function
for modelling of the loss level:
200
5. APPLICATION IN DIALOGUES – EPS SIMULATION 5.4.2
Outage length Loss function
(t, in minutes) ($)
0–5 30 000 + (t ∗ 1 000)
5–20 35 000 + (t ∗ 700)
> 20 49 000 + (t ∗ 200)
The factory is running overnight, all time with approximately the same
power consumption (we can consider a consumption of non-machinery
parts of the factory insignificant). The average value is 0.5 MW again
with Gaussian changes with the mean of 3 kW and the deviation of 0.1.
The agent G is the only facility in the system which has two inputs.
Therefore it is the only agent for which we must supply a special implementation
of the reaction to input outages. The behaviour will be
straightforward: to switch to another power supply line. In our first run
of the simulation, the switch time (time to recover from the outage) will
be 3 minutes and in the second run 10 minutes.
The last consumer point is the agent F which represents a city quarter.
Hypothetical losses of this consumer are constant in different periods of
time:
Daytime Loss function
(hours) ($)
0am–8am $100 per hour
8am–5pm $300 per hour
5pm–12pm $500 per hour
The average power consumption of the agent F is 50 kW in the morning
and 200 kW in the afternoon period of the day. The variation of the
Gaussian changes is the constant of 3 kW.
5.4.2 The Agents Implementation
In our example, we use standard agent features whose templates are
supplied by the Rice multi-agent framework libraries. In the agents implementation
we can use standard functions and instantiate the agents
classes. Rice supplies a function which runs each agents in its own thread.
All functions in Rice are thread-safe.
Power lines are fault-free so we can simply instantiate standard class
implementation:
201
5.4.2 5. APPLICATION IN DIALOGUES – EPS SIMULATION
Line1=BaseLine(1, timer, HelperInterface.A_RL_LINE,
"L1", 0, "A", "B")
For other agents we will create corresponding classes which model particular
agent behaviour. Rice uses event-driven model which means that
a “hook” (trigger) function can be defined for each event processed by
the agent. There will be only a few functions which are to be redefined.
Some of them will be illustrated here (grouped by the behaviour class as
stated above).
5.4.2.1 Event-Based Behaviour
Demand initialization – consumer agents (E, F, G, H) set their initial
power demand value to the mean of their power consumption.
class BaseConsumer(BaseFacility):
def __init__(self, demandMean, demandVar,...)
self.initPowerDemand = demandMean
self.demandMean = demandMean
self.demandVariation = demandVar
Reaction to outage – all agents (except G) will set itself into the outage
state if the input lines are in the outage state, too. This is the standard
behaviour. For the agent G a different action is implemented – try to
switch its inputs:
def outageBeginHook(self, cause):
for input in inputs:
res = input.agent.powerDemand(self.lastDemand)
if not res==0:
# power granted, status remains the same
return
# no power granted
self.setRLStatus(StatusOutage(cause))
5.4.2.2 Idle-Loop Behaviour
We must continuously check the agent state and generate errors in particular
time intervals. Also we want to simulate continuous changes of the
202
5. APPLICATION IN DIALOGUES – EPS SIMULATION 5.4.2
Figure 5.14: Power flow graph of part of the simulation of 10 years period
produced by Rice.
power consumption. This will be performed by implementation of two
hooks which are present in the standard Rice idle-loop implementation:
demandChangeHook and statusChangeHook.
We will use standard Gaussian distribution of power consumption
value changes. The corresponding function is supplied by the Rice Utils
library.
def demandChangeHook(self, cause):
return rice.Utils.gaussian(self.demandMean,
self.demandVariation)
To simulate the probability of facility errors we will use another standard
function. Duration of the errors has also Gaussian distribution. The end
of the error will be triggered by the timer.
def statusChangeHook(self, cause):
if isinstance(self.rlStatus, StatusOK):
res = rice.Utils.elapsed(self.timer,
self.errorProbab,
203
5.4.3 5. APPLICATION IN DIALOGUES – EPS SIMULATION
self.errorVar)
if not res==0:
self.setRLStatus(StatusError(cause=self.aid))
self.setTimer(
rice.Utils.gaussian(self.errorDurMean,
self.errorDurVar))
# ...
These snippets of the example code illustrate that the implementation of
a particular power system model within the Rice system is straightfor-
ward.8
5.4.3 The Simulation Run
The example simulation of the EPS used for computation of economic
losses ran for ten years of the simulated time. The actual running time
was 8 minutes. An example graph of power flow on the inputs of the
agents in one period of the simulation can be seen in the Figure 5.14.
The resulting numbers and durations of power supply outages and
the computed economic losses are summarized in the Table 5.5. Errors
in the power supply from agent D cause only losses for agent F. G has
two inputs, B and C. The initial input is C. When there is an error
in the power delivery from C, the input is switched to B (which has a
limited capacity). In the simulation, we have measured and computed
the economic impacts of the two situations denoted as G1 (without using
the B backup input) and G2 (with both inputs switching). Even if there
are two inputs, the power supply would not be entirely uninterrupted
because of the B’s supply limitation. The power restoration of G depends
on the actual power consumption of agents E and F. We can see that
the provision of another input for G has saved $445 125. The time when
G was switched to input B was too short to cause any overloading on
B (denying of the demands of E and F). The agent H is not active for
the whole day, which is why only 3 of its outages have caused an overall
economic loss of $1 183.
The nature of the simulation is strictly probabilistic. The resulting
values vary in different runs of the simulation. In this one-run simulation
the result was that the possible overall losses caused by not supplied
8The whole implementation of the example has about 200 lines of code in Python.
204
5. APPLICATION IN DIALOGUES – EPS SIMULATION 5.4.3
Table 5.5: The outages and economic loss during the simulation run
Agent Outages time Economic loss
minutes
A 0 →B, →C
B 0 →E, →D, →G
C 213 →G, →H
D 7 970 →F
E 262 $13 110
F (total) 7 970 $42 416
F: 0am–8am 2 593 $4 321
F: 8am–17pm 2 015 $10 076
F: 17pm–24pm 3 362 $28 019
G1 (one input) 213 $555 321
G1: 0–5min out. 5 $34 510
G1: 5–20min out. 34 $93 961
G1: 20–∞min out. 174 $426 850
G2 (two inputs) 47 $110 196
G2: 0–5min out. 0 $0
G2: 5–20min out. 20 $48 818
G2: 20–∞min out. 27 $61 378
H 213 $1 183
energy will be $612 030 if the backup power line from B to G was not
provided and only $166905 if it was. We can see clearly how big the
financial impact of an existence of such line is.
The number of intervening agents, events and actions in a real power
system simulation is obviously higher than in the presented example with
higher number of possible dynamic changes of the network topology. But
we believe that the design of the Rice system allows us to cope with any
such modelled situation – the next step lies in testing the system on a part
of a real power network from the Ostrava region in the Czech Republic.
205
5.5 5. APPLICATION IN DIALOGUES – EPS SIMULATION
5.5 Future Work on the Rice System
We have described the design and prototype implementation of a multiagent
system for active simulation of energy flow in an electrical power
network. The implementation framework meets all the required criteria
which were identified in the process of power systems simulation analysis.
It has already proven to be enough flexible, open, dynamic and robust
even for possible real-time power systems monitoring.
The Rice system for simulation of power network processes provides a
modular platform for analysis of the energy flows within any large-scale
power system. During the design of the project, the system concentrates
on handling of uncertainty either in the form of approximate values from
the network or unknown behaviour of the take-off points.
The new assets that Rice brings to the field of power system simulation
lie in the estimated calculation of the economic aspects of the power
system outage, which is enabled by the modular design of the multi-agent
network.
We have also described the first stages of implementing a natural language
dialogue system for human-computer communication specialized
to the domain of electrical power network field. The main purpose of the
designed system is to support complicated queries of large time SCADA
databases of electrical network failures included in a system for prediction
and prevention of blackout and power network failures.
The text includes the description of building a corpus of electrical
power networks texts as well as preliminary results of natural language
processing analysis of the texts. These results show the need of adapting
the current NL tools for the purpose of analysis of such specific domain.
Extending the tools forms the baseline of our future directions on this
topic of research.
The future directions of the implementation of the system lie mainly in
testing the system with real-world data (which are actually very difficult
to obtain due to their economic meaning). Nevertheless, we believe that
the system can form an important acquisition with the ability to test and
predict possible system outages of power networks.
206
Chapter 6
Conclusions and Future
Directions
We have presented the current results of several research projects implemented
in the NLP Centre at the Faculty of Informatics, Masaryk
University. The central topics of all these projects are the syntactic and
semantic analysis of natural language sentences with the concentration
on the Czech language. The projects are in different stages of development,
but all of them have already proved that they can significantly
contribute to the field of computer processing of natural languages.
The Czech valency lexicon VerbaLex is still in its development phase,
however, we can see its potential exploitation in large number of practical
natural language processing applications, not only the syntactic parsing,
but also problems like word sense disambiguation, predicate-argument
structure extraction, machine translation or logical analysis. The linkage
to the Princeton WordNet synsets provides the possibility of multilingual
projections of VerbaLex to other languages. The experiments
with using Czech valency frames for building Bulgarian and Romanian
frames in Balkanet project have been evaluated as more than promising,
see [TMBM06] and S. Koeva in [Chr04].
The usage of electronic dictionary resources in the computational linguistics
has increased several times during the last decade due to evident
reasons: 1) the computer equipment has reached such high state that it
207
6. CONCLUSIONS AND FUTURE DIRECTIONS
is feasible to process giga bytes of textual data [Gra03], and 2) the effectiveness
gain is enormous when compared to the previous types of slow
“manual” processing of linguistic information.
Despite this fact, the list of generally available dictionary writing
systems (DWS) is not very wide – we may refer to e.g. Longman Dictionary
Publishing System [McN03], TshwaneLex [JdS04] or the Dictionary
Editor and Browser (DEB) [HPRR06, HP06a]. All the cited systems
are based on XML databases that allow to capture practically any kind
of structural data including monolingual and translational dictionaries,
thesauri or encyclopediae. Longman DPS and TshwaneLex are self contained
commercial applications that are designed for specific purposes –
Longman DPS is used in the publishing house to bring out traditional
paper based dictionaries as well as new electronic and on-line products.
TshwaneLex on the other hand is a dictionary compilation system that
allows to create and maintain several dictionary styles for various purposes.
Both these systems are distributed on a commercial base.
The third system, the DEB platform, is an open source and freely
available development framework developed at the Centre of Natural
language processing at Masaryk University, Czech Republic. The system
provides strict client server architecture for design of completely versatile
dictionary applications. We have described the details of the system as
a whole including the presentation of a new component, the Administration
interface that is used by all client applications. The relevant features
of the DEB platform are high modularity, configurability and flexibility
which results in an easy adaptability for the various tasks. Thanks to
them, the DEB platform represents a versatile base, on which the individual
and powerful dictionary writing tools (clients) such as DEBVisDic
are implemented.
In our view the XML formats within the DEB platform present a
reasonable base not only for merging various lexical resources but also
for their future standardization. This has been convincingly shown in the
Balkanet Project where the VisDic XML format has been employed for
building all 6 Balkanet languages (Bulgarian, Czech, Greek, Romanian,
Serbian and Turkish) plus English. In fact, VisDic tool played the role
of the instrument through which the first WordNet standardization steps
have started. It is our belief that DEBVisDic tool can play the same role
in the near future as well.
The DEB platform is being (and will be) thoroughly tested with its
208
6. CONCLUSIONS AND FUTURE DIRECTIONS
clients used in several teams all over the world. We may name e.g. the
Dutch Cornetto project or the Polish and Hungarian WordNet projects.
We also discuss the possibility of DEBVisDic being used as the main
WordNet tool in the near future as well, namely in the preparation of the
Global WordNet Grid.
The PRALED client is used in the Institute of Czech Language, Czech
Academy of Sciences (Prague) as a dictionary writing system for building
the Czech Lexical Database which is a large project planned for about
5 more years from now. The goal is to develop a lexical database of
contemporary Czech containing approximately 100 000 entries.
Full parsing of natural language sentences is a complicated task that
provides a transition from unstructured text into structural information
suitable for all forms of information retrieval. Any kind of higher level
language understanding and/or semantic processing must rely on the
results of syntactic parsing. The quality of state-of-the-art syntactic
parsers [Cha01, Col99] is still not completely satisfiable even for analytical
languages like English. In case of free-word-order languages like
Czech, the situation is even more complicated.
At the Centre of Natural Language Processing at the Faculty of Informatics,
Masaryk University in Brno, a full syntactic parsing system is
being developed since 1997. The parsing system synt currently provides
fast and reliable algorithm that provides basis for higher-level structural
processing of common natural language texts. Nowadays, the methods
for best analysis selection are being implemented in the system – some
of these methods were in detail presented in this text. We believe that
by improving the synt tree ranking algorithm with these methods, we
will have one of the best Czech parsers at hand for the needs of the
Normal Translation Algorithm for automatic semantic analysis of Czech
sentences by means of the Transparent Intensional Logic (TIL).
Pavel Tichý published his views on the language analysis in TIL in
several papers (e.g. [Tic94a]) and started to write a thorough work on that
matter [Tic94b]. However, he managed to write only the first out of the
intended twelve chapters, and thus has left a lot of particular phenomena
of natural language without the prescription of their proper analysis. The
first text, that tries to describe the algorithm of translation of natural
language sentence in its completeness, is the cited work [Hor02]. Nevertheless,
the algorithm is still in its first version only and its finalization is
the work of Tichý’s followers such as Materna, Oddie, Duží, Horák and
209
6. CONCLUSIONS AND FUTURE DIRECTIONS
others.
The sentence analysis using TIL in our work is directed not only to
the mere meaning representation, but also provides a systematic basic for
automatic logical inference of new facts. Tichý has proposed a theoretical
framework for the TIL inference in [Tic82], which is further discussed
and elaborated in [DJM08]. In a long term NLP Centre project, we are
developing particular methods for a question-answering mechanism based
on knowledge base data structures from [Hor02, chapter 6]. The first
prototype implementation of such system have already been published
in [GH07a] and [GH07b].
In the last chapter, we have presented the Rice multi-agent system
that allows working with the power network processes without the need of
any of the network component but still with the power to define and capture
any complex behaviour of the components in real situations including
the possibility of automatic estimation of the economic loss caused by a
network failure.
The described communication framework follows all the required criteria
which were identified in the process of power systems simulation
analysis. It has already proven to be flexible enough, open, dynamic and
robust even for possible real-time power systems monitoring. This is continuously
validated by the growing implementation and it will be tested
in cooperation with large Czech power supply companies. The system
architecture is suitable for usage in largely decentralized networks of autonomous
data gathering and processing units. Especially an integration
with networks of thin-client sensors (e.g. ZigBee [Zig05]) deserves further
attention because of the promising industrial applications.
Currently, the system exploits the valuable databases of power system
failures which have been constructed for several years in 9 regions of the
Czech Republic and Slovakia. Further development aims at integrating
the results in the stochastic simulation of failures of particular network
elements which, as we hope, will be an important step towards the tools
and systems helping to effectively control the complex behavior of such
emergent systems.
210
Bibliography
[Aea93] E. Abel et al. A Multiagent Approach to Analyze Disturbances in
Electrical Networks. In Proceedings of the Fourth ESAP Conference,
pages 606–611, Melbourne, Australia, 1993. 160
[AH02] J. Aycock and R.N. Horspool. Practical Earley Parsing. The
Computer Journal, 45(6):620–630, 2002. 86
[Ami01] M. Amin. Toward self-healing energy infrastructure systems. the
IEEE Computer Applications in Power journal, 14, part 1:20–28,
2001. 162
[BA96] R. Billinton and R.N. Allan. Reliability Evaluation of Power Systems.
Plenum Press, New York, second edition, 1996. 196
[Bal04] The Balkanet Project Website. http://www.ceid.upatras.gr/
Balkanet/, 2004. 8, 10
[BBHS+
04] S. Baumann, C. Brinckmann, S. Hansen-Schirra, et al. The MULI
Project: Annotation and Analysis of Information Structure in German
and English. In Proceedings of the LREC 2004 Conference,
Lisboa, Portugal, 2004. 69
[BBR87] G. E. Barton, R. C. Berwick, and E. S. Ristad. Computational
complexity and natural language. MIT Press, Cambridge, Massachusetts,
1987. 85
[BKPR06] M. Baroni, A. Killgarriff, J. Pomikálek, and P. Rychlý. Webbootcat:
a web tool for instant corpora. In Proceeding of the EuraLex
Conference 2006, pages 123–132, Italy, 2006. Edizioni dell’Orso
s.r.l. 180
[BN00] H. Bunt and A. Nijholt, editors. Advances in Probabilistic and
Other Parsing Technologies. Kluwer Academic Publishers, 2000.
70, 90
211
BIBLIOGRAPHY BIBLIOGRAPHY
[Bod03] Rens Bod. An efficient implementation of a new DOP model. In
EACL 2003, pages 19–26, 2003. 76
[BPGR06] Hans C. Boas, Elias Ponvert, Mario Guajardo, and Sumeet Rao.
The current status of German FrameNet. In SALSA workshop at
the University of the Saarland, Saarbrucken, Germany, 2006. 14
[BS05] Gregory Burd and Kimbro Staken. Use a Native XML Database
for Your XML Data. XML Journal, May. 26, 2005, 2005. http://
xml.sys-con.com/read/90126.htm. 40
[BSDH98] Srinivas Bangalore, Anoop Sarkar, Christine Doran, and Beth Ann
Hockey. Grammar & Parser Evaluation in the XTAG Project,
1998. http://www.cs.sfu.ca/~anoop/papers/pdf/eval-final.
pdf. 115
[BSS98] A. Barceinas, J. A. Sánchez, and J. L. Schnase. MICK: A
KQML inter-agent communication framework in a digital library.
In Memorias del Simposium Internacional de Computación
(CIC’98), Mexico City, 1998. 167
[C+
04] K. Czajkowski et al. The WS-Resource Framework, 2004. http://
www.globus.org/wsrf/specs/ws-wsrf.pdf. 160
[CD02] Michael Collins and Nigel Duffy. New Ranking Algorithms for
Parsing and Tagging: Kernels over Discrete Structures, and the
Voted Perceptron. In ACL 2002, pages 263–270, 2002. 76
[CJa02] Object Management Group. IDL to Java Language Mapping Specification.,
2002. http://www.omg.org/technology/documents/
formal/omg_idl_to_java_language_mapping.htm. 173
[CNC06] The Czech National Corpus, 2006. http://ucnk.ff.cuni.cz/
english/. 7
[COB04] Institut de Recherche en Informatique de Toulouse. The COBALT
project, 2004. http://www.irit.fr/recherches/SIERA/GRS/
coop.frame.shtml. 167
[Col98] Michael Collins. dep2phr – conversion between dependency
and phrase structures, 1998. http://ufal.mff.cuni.cz/pdt/
Utilities/dep2phr/. 110, 112
[Col99] M. Collins. Head-driven Statistical Models for Natural Language
Parsing. PhD thesis, Computer Science Department, University
of Pennsylvenia, Philadelphia, 1999. 209
[COR04] Object Management Group. Common Object Request Broker
Architecture Core Specification., 2004. http://www.omg.org/
technology/documents/formal/corba_iiop.htm. 173
212
BIBLIOGRAPHY BIBLIOGRAPHY
[CPy02] Object Management Group. Python Language Mapping Specification.,
2002. http://www.omg.org/technology/documents/
formal/python.htm. 173
[Cre75] M.J. Cresswell. Hyperintensional Logic. Studia Logica, (34):25–38,
1975. 123
[Cre85] M.J. Cresswell. Structured meanings. MIT Press, Cambridge,
Mass, 1985. 123
[Cro06] Douglas Crockford. JSON, The Fat-Free Alternative to XML. In
Proceedings of XML 2006, Boston, USA, 2006. http://www.json.
org/xml.html. 40
[CRZ03] Akmal B. Chaudhri, Awais Rashid, and Roberto Zicari. XML Data
Management: native XML and XML-enabled database systems.
Addison-Wesley Professional, 2003. http://books.google.com/
books?id=7LNhdOeQulQC. 40
[CVJ92] D. Cockburn, L. Z. Varga, and N. R. Jennings. Cooperating Intelligent
Systems for Electricity Distribution. In Proceedings of BCS
Expert Systems 92 Conference (Application Track), Cambridge,
UK, 1992. 167
[Č+
83] F. Čermák et al. Slovník české frazeologie a idiomatiky I-IV
(Dictionary of Czech Phraseology and Idioms, SČFI). Academia,
Praha, 1983. 57, 61
[DBX07] Oracle Berkeley DB XML, 2007. http://www.oracle.com/
database/berkeley-db/xml. 40
[Deb03] R. Debusmann. A Parser System for Extensible Dependency
Grammar, pages 103–106. Loria, Nancy France, 2003. 71
[DJM08] Marie Duží, Bjorn Jespersen, and Pavel Materna. Transparent Intensional
Logic. Foundations and Applications. Springer Science,
the Netherlands, 2008. to appear. 121, 210
[DKPR98] Hoa Trang Dang, Karin Kipper, Martha Palmer, and Joseph
Rosenzweig. Investigating regular sense extensions based on intersective
Levin classes. In Proceedings of Coling-ACL98, Montreal
CA, August 11-17, 1998. www.cis.upenn.edu/~mpalmer/. 13
[DTH99] L. DiPippo, B. Thuraisingham, and E. Hodys. Towards a RealTime
Agent Architecture – A Whitepaper. Technical report, 1999.
TR99-273. 167
[Ear70] J. Earley. An efficient context-free parsing algorithm. In Communications
of the ACM, volume 13, pages 94–102, 1970. 83
213
BIBLIOGRAPHY BIBLIOGRAPHY
[Erl04] Jens Erlandsen. iLEX – an ergonomic and powerful tool combining,
effective and flexible editing with easy and fast search and
retrieval. In EURALEX 2004, Lorient, France, 2004. demonstration.
9
[Eur99] The EuroWordNet Project Website. http://www.illc.uva.nl/
EuroWordNet/, 1999. 8
[F+
95] J. Filipec et al. Slovník spisovné češtiny (Dictionary of Literary
Czech, SSČ). Academia, Praha, 1st
edition, 1995. electronic version,
LEDA, Praha. 56, 61
[FBS04] C.J. Fillmore, C.F. Baker, and H. Sato. FrameNet as a ’net’.
In Proceedings of Language Resources and Evaluation Conference
(LREC 04), volume vol. 4, 1091-1094, Lisbon, 2004. ELRA. 14,
61
[FFMM94] T. Finin, R. Fritzson, D. McKay, and R. McEntire. KQML as
an Agent Communication Language. In Proc. of 3rd
Int. Conf.
On Information and Knowledge Management CIKM 1999. ACM
Press, 1994. 174
[Fis99] D. A. Fisher. Design and Implementation of EASEL. In Proceedings
of MacHack 14, the 14th
Annual Conference for Leading Edge
Developers, Deeerborn, MI, 1999. 5, 153
[GAL03] Pablo Gamallo, Alexandre Agustini, and Gabriel P. Lopes. Learning
Subcategorisation Information to Model a Grammar with Corestrictions.
Traitement Automatique de la Langue, 44(1):93–117,
2003. 91
[Gar07] A. Gardoň. The Object layer of the Dolphin database. http://
nlp.fi.muni.cz/projects/dolphin/object_layer.htm, 2007.
146
[GF92] Michael R. Genesereth and Richard E. Fikes, editors. Knowledge
Interchange Format, Version 3.0, Reference Manual. Stanford
University, Stanford CA, USA, 1992. http://www-ksl.stanford.
edu/knowledge-sharing/kif/. 176
[GH07a] Andrej Gardoň and Aleš Horák. Dolphin – a Knowledge Base
for Transparent Intensional Logic. In Proceedings of the Seventh
International Workshop on Computational Semantics (IWCS-7),
Tilburg, The Netherlands, 2007. 120, 210
[GH07b] Andrej Gardoň and Aleš Horák. The Learning and Question Answering
Modes in the Dolphin System for the Transparent Intensional
Logic. In Proceedings of the First Workshop on Recent Ad-
214
BIBLIOGRAPHY BIBLIOGRAPHY
vances in Slavonic Natural Language Processing (RASLAN 2007),
Brno, Czech Republic, 2007. 210
[Gra03] David Graff. English Gigaword. Technical Report LDC2003T05,
Philadelphia, PA USA, 2003. 208
[GRA07] GRASS – Geographic Resources Analysis Support System, 2007.
http://grass.itc.it/. 42
[GWA07] The Global WordNet Association. http://www.globalwordnet.
org/, 2007. 9
[Haj04a] Jan Hajič. Complex Corpus Annotation: The Prague Dependency
Treebank. Bratislava, Slovakia, 2004. Jazykovedný ústav Ľ. Štúra,
SAV. 7, 87, 107, 110, 157
[Haj04b] Jan Hajič. Disambiguation of Rich Inflection (Computational Morphology
of Czech). Karolinum, Charles University Press, Prague,
Czech Republic, 2004. 109
[Haj98] J. Hajič. Building a Syntactically Annotated Corpus: The Prague
Dependency Treebank. In Issues of Valency and Meaning, pages
106–132, Prague, 1998. Karolinum. 75
[Han04] Patrick Hanks. Corpus Pattern Analysis. In Proceedings of
the Eleventh EURALEX International Congress, Lorient, France,
2004. Universite de Bretagne-Sud. 14, 59
[Hav57] Bohumil Havránek, editor. Příruční slovník jazyka českého (Reference
Dictionary of Czech Language, PSJČ). Státní nakladatelství/SPN,
Praha, 1933–1957. 56
[HCRT99] Jan Hajič, Michael Collins, Lance Ramshaw, and Christoph Tillmann.
A Statistical Parser for Czech. In Proceedings ACL’99,
Maryland, USA, 1999. 108
[HH05a] Dana Hlaváčková and Aleš Horák. VerbaLex – New Comprehensive
Lexicon of Verb Valencies for Czech. In Proceedings of the
Computer Treatment of Slavic and East European Languages 2005,
pages 107–115, Bratislava, Slovakia, 2005. 10
[HH05b] Aleš Horák and Dana Hlaváčková. Transformation of WordNet
Czech Valency Frames into Augmented VALLEX-1.0 Format. In
Proceedings of LTC 2005, Poznan, Poland, 2005. 10
[HH07] J. Hajič and E. Hajičová. Some of Our Best Friends Are Statisticians.
Lecture Notes in Computer Science, 4629:2–10, 2007. 3
[HHK06] Dana Hlaváčková, Aleš Horák, and Vladimír Kadlec. Exploitation
of the VerbaLex Verb Valency Lexicon in the Syntactic Analysis
215
BIBLIOGRAPHY BIBLIOGRAPHY
of Czech. In Proceedings of Text, Speech and Dialogue 2006, pages
79–85, Brno, Czech Republic, 2006. Springer-Verlag. 71
[HHKK07] Aleš Horák, Tomáš Holan, Vladimír Kadlec, and Vojtěch Kovář.
Dependency and Phrasal Parsers of the Czech Language: A Comparison.
In Proceedings of the 10th
International Conference on
Text, Speech and Dialogue, Pilsen, Czech Republic, 2007. accepted
for publication. 71, 106
[HK05a] Aleš Horák and Vladimír Kadlec. New Meta-grammar Constructs
in Czech Language Parser synt. In Proceedings of Text, Speech
and Dialogue 2005, pages 85–92, Karlovy Vary, Czech Republic,
2005. Springer-Verlag. 70, 71
[HK05b] Aleš Horák and Vladimír Kadlec. Czech Language Parsing using
Meta-grammar Formalism with Contextual Constraints. In Proceedings
of the Computer Treatment of Slavic and East European
Languages 2005, pages 124–129, Bratislava, Slovakia, 2005. 71
[HK06] Aleš Horák and Vladimír Kadlec. Platform for Full-Syntax Grammar
Development Using Meta-grammar Constructs. In Proceedings
of the 20th
Pacific Asia Conference on Language, Information
and Computation, pages 311–318, Beijing, China, 2006. Tsinghua
University Press. 71
[HKS02] A. Horák, V. Kadlec, and P. Smrž. Enhancing Best Analysis Selection
and Parser Comparison. In Lecture Notes in Artificial Intelligence,
Proceedings of TSD 2002, pages 461–467, Brno, Czech
Republic, 2002. Springer Verlag. 71, 106
[HN05] K. Hall and V. Novák. Corrective modeling for non-projective
dependency parsing. pages 42–51, 2005. 108, 111
[Hof95] B. Hoffman. The Computational Analysis of the Syntax and Interpretation
of Free Word Order in Turkish. PhD thesis, University
of Pennsylvania, Philadelphia, 1995. 69
[Hol04] Tomáš Holan. Tvorba závislostního syntaktického analyzátoru
(Building a dependency syntactic analyser). In Sborník semináře
MIS 2004. Matfyzpress, Prague, Czech Republic, 2004. In Czech.
108
[Hol05] Tomáš Holan. Genetické učení závislostních analyzátorů (Genetic
learning of dependency analysers). In Sborník semináře ITAT
2005. UPJŠ, Košice, 2005. In Czech. 108
[Hor02] Aleš Horák. The Normal Translation Algorithm in Transparent
Intensional Logic for Czech. PhD thesis, Faculty of Informatics,
216
BIBLIOGRAPHY BIBLIOGRAPHY
Masaryk University, Brno, 2002. iv, 4, 69, 119, 120, 121, 124, 134,
136, 140, 144, 150, 155, 209, 210
[Hor04] Aleš Horák. Types in Transparent Intensional Logic and Easel – a
Comparison. In Proceedings of the IASTED International Conference
Artificial Intelligence and Applications 2004, pages 833–837,
Anaheim, Calgary, Zurich, 2004. The International Association of
Science and Technology for Development. 120
[HP04] Aleš Horák and Karel Pala. Lexicons in TIL and Verb Valency
Frames. In Proceedings of the International Conference on Communications
in Computing (CIC 2004), pages 255–261, Las Vegas,
Nevada, USA, 2004. CSREA Press. 120
[HP06a] Aleš Horák and Karel Pala. DEB tools for merging linguistic resources.
In Proceedings of the Workshop on Layering Linguistic
Information, LREC 2006, pages 55–61, Italy, 2006. ELRA. 10,
208
[HP06b] A. Horák and M. Prýmek. Modelling of Economical Aspects of a
Power System Failure. In the Proceedings of the International
Conference on Power, Energy, and Applications (PEA 2006),
Gaborone, Botswana, 2006. 162
[HP06c] A. Horák and M. Prýmek. State-oriented Maintenance of Electrical
Power Systems with Dynamic Multi-agent Network. In Proceedings
of CSIT 2006, pages 310–316, Amman, Jordan, 2006. ASP
University. 162
[HP07] Aleš Horák and Karel Pala. Building a Large Lexicon of Complex
Valency Frames. In Proceedings of the FRAME 2007: Building
Frame Semantics Resources for Scandinavian and Baltic Languages,
pages 31–38, Tartu, Estonia, 2007. Lund University, Sweden.
10
[HPDM07] Aleš Horák, Karel Pala, Marie Duží, and Pavel Materna. Verb Valency
Semantic Representation for Deep Linguistic Processing. In
Proceedings of the Workshop on Deep Linguistic Processing, ACL
2007, pages 97–104, Prague, Czech Republic, 2007. the Association
for Computational Linguistics. 120
[HPRP06] Aleš Horák, Karel Pala, Adam Rambousek, and Martin Povolný.
First Version of New Client-Server WordNet Browsing and Editing
Tool. In Proceedings of the Third International WordNet Conference
- GWC 2006, pages 325–328, Jeju, South Korea, 2006.
Masaryk University, Brno. 10
217
BIBLIOGRAPHY BIBLIOGRAPHY
[HPRR06] Aleš Horák, Karel Pala, Adam Rambousek, and Pavel Rychlý. New
clients for dictionary writing on the DEB platform. In DWS 2006:
Proceedings of the Fourth International Workshop on Dictionary
Writings Systems, pages 17–23, Italy, 2006. Lexical Computing
Ltd., U.K. 10, 208
[HR07] Aleš Horák and Adam Rambousek. Dictionary Management System
for the DEB Development Platform. In Proceedings of the
4th
International Workshop on Natural Language Processing and
Cognitive Science (NLPCS, aka NLUCS), pages 129–138, Funchal,
Portugal, 2007. INSTICC PRESS. 10
[HS00] A. Horák and P. Smrž. Large Scale Parsing of Czech. In Proceedings
of Efficiency in Large-Scale Parsing Systems Workshop,
COLING’2000, pages 43–50, Saarbrücken, 2000. Universität des
Saarlandes. 77, 82, 85
[HS02] Aleš Horák and Pavel Smrž. Best Analysis Selection in Inflectional
Languages. In Proceedings of the 19th international conference on
Computational linguistics, pages 363–368, Taipei, Taiwan, 2002.
Association for Computational Linguistics. 69, 90, 110
[HS03] Aleš Horák and Pavel Smrž. VisDic – WordNet Browsing and
Editing Tool. In Proceedings of the Second International WordNet
Conference – GWC 2004, pages 136–141, Brno, Czech Republic,
2003. 10, 25
[HS04] Aleš Horák and Pavel Smrž. New Features of WordNet Editor VisDic.
In Romanian Journal of Information Science and Technology,
volume 7(1–2), pages 1–13, 2004. 10, 39
[HVR08] Aleš Horák, Piek Vossen, and Adam Rambousek. A Distributed
Database System for Developing Ontological and Lexical Resources
in Harmony. In Lecture Notes in Computer Science, Proceedings
of CICLing 2008, Haifa, Israel, 2008. Springer-Verlag. 10
[HŽ06] Tomáš Holan and Zdeněk Žabokrtský. Combining Czech Dependency
Parsers. In Lecture Notes in Artificial Intelligence, Proceedings
of TSD 2006, pages 95–102, Brno, Czech Republic, 2006.
Springer Verlag. 108
[Cha01] E. Charniak. Immediate-head parsing for language models. In Proceedings
of the 39th Annual Meeting of the Association of Computational
Linguistics, Toulouse, France, 2001. 209
[Chr04] D. Christodoulakis. Balkanet Final Report. University of Patras,
DBLAB, 2004. No. IST-2000-29388. 207
218
BIBLIOGRAPHY BIBLIOGRAPHY
[ICU07] International Components for Unicode (ICU), 2007. http://ibm.
com/software/globalization/icu, Open-source C/C++ and
Java libraries for Unicode and I18N support. 42
[IDL02] Object Management Group. OMG IDL Syntax and Semantics.,
2002. http://www.omg.org/cgi-bin/doc?formal/02-06-07. 173
[JB03] N. R. Jennings and S. Bussmann. Agent-based control systems.
Number 23(3), pages 61–74, 2003. 168
[JdS04] D. Joffe and G-M. de Schryver. TshwaneLex – Professional off-theshelf
lexicography software. In Third International Workshop on
Dictionary Writing Systems, Brno, Czech Republic, 2004. Masaryk
University, Faculty of Informatics. 9, 208
[Jen00] N. R. Jennings. On Agent-Based Software Engineering. Artificial
Intelligence, 17(2):277–296, 2000. 162, 168
[JG02] T.F. Jaeger and V.A. Gerassimova. Bulgarian word order and
the role of the direct object clitic in LFG. In M. Butt and T.H.
King, editors, Proceedings of the LFG02 Conference, Stanford,
2002. CSLI Publications. 69
[Kay89] M. Kay. Algorithm Schemata and Data Structures in Syntactic
Processing. In Report CSL-80-12, Palo Alto, California, 1989.
Xerox PARC. 83
[KDP00] Karin Kipper, Hoa Trang Dang, and Martha Palmer. Class Based
Construction of a Verb Lexicon. In AAAI-2000 Seventeenth National
Conference on Artificial Intelligence, Austin TX, 2000. 14
[KH07] Vojtěch Kovář and Aleš Horák. Reducing the Number of Resulting
Parsing Trees for the Czech Language Using the Beautified Chart
Method. In Proceedings of the 3rd
Language & Technology Conference:
Human Language Technologies as a Challenge for Computer
Science and Linguistics, Poznan, Poland, 2007. accepted for publication.
71
[KKH06] Vojtěch Kovář, Vladimír Kadlec, and Aleš Horák. Grammar Development
for Czech Syntactic Parser with Corpus-based Techniques.
In Proceedings of Corpus Linguistic 2006, pages 159–165,
Saint-Petersburg, Russia, 2006. Saint-Petersburg State University.
71
[KKK00] J. Kocek, M. Kopřivová, and K. Kučera, editors. Český národní
korpus – úvod a příručka uživatele (The Czech National Corpus –
introduction and user’s handbook). ÚČNK FF UK, Prague, 2000.
7
219
BIBLIOGRAPHY BIBLIOGRAPHY
[KKRP06] Karin Kipper, Anna Korhonen, Neville Ryant, and Martha
Palmer. Extensive Classifications of English verbs. In Proceedings
of the 12th
EURALEX International Congress, Turin, Italy,
2006. 134
[KMG01] A.K. Khan, R.P. Meters, and C.A. Los Gatos. Monitoring power
for the future. Power Engineering Journal, 15(2):81–85, Apr 2001.
179
[KP+
99] J. Kraus, V. Petráčková, et al. Akademický slovník cizích slov
(Academic Dictionary of Foreign Words, SCS). Academia, Praha,
1999. electronic version, LEDA, Praha. 56, 61
[KRST04] Adam Kilgarriff, Pavel Rychlý, Pavel Smrž, and David Tugwell.
The Sketch Engine. In Proceedings of the Eleventh EURALEX International
Congress, pages 105–116, Lorient, France, 2004. Universite
de Bretagne-Sud. 61, 135
[Kru03] G. Kruijff. 3-phase grammar learning. In Proceedings of Workshop
on Ideas and Strategies for Multilingual Grammar Development,
Vienna Austria, 2003. 71
[KS06] Vladimír Kadlec and Pavel Smrž. How Many Dots Are Really
Needed for Head-Driven Chart Parsing? In Proceedings of SOFSEM
2006, pages 483–492, Czech Republic, 2006. Springer-Verlag.
83
[Lev93] Beth Levin, editor. English Verb Classes and Alternations: A Preliminary
Investigation. The University of Chicago Press, Chicago,
1993. 13
[Lou98] M. Louw. Polaris User’s Guide. Technical report, Lernout & Hauspie
– Antwerp, Belgium, 1998. 8
[Mat04] P. Materna. Conceptual Systems. Logos Verlag, Berlin, 2004. 121
[Mat98] P. Materna. Concepts and Objects, volume 63 of Acta Philosophica
Fennica. The Philosophical Society of Finland, Helsinki, 1998. 150
[MC07] Robert C. Moore and John Carroll. Parser Comparison – ContextFree
Grammar (CFG) Data, 2007. http://www.cogs.susx.ac.
uk/lab/nlp/carroll/cfg-resources/. 106
[McD06] Ryan McDonald. Discriminative learning and spanning tree algorithms
for dependency parsing. PhD thesis, University of Pennsylvania,
2006. 108
[McN03] Michael McNamara. Dictionaries for all: XML to Final Product.
In XML Conference 2003, Philadelphia, USA, 2003. 9, 208
220
BIBLIOGRAPHY BIBLIOGRAPHY
[Mil90] G. Miller. Five Papers on WordNet. International Journal of
Lexicography, 3(4), 1990. Special Issue. 8, 129
[MK91] J. T. Maxwell III and R. M. Kaplan. The Interface between
Phrasal and Functional Constraints. In M. Rosner, C. J. Rupp, and
R. Johnson, editors, Proceedings of the Workshop on Constraint
Propagation, Linguistic Description, and Computation, pages 105–
120. Instituto Dalle Molle IDSIA, Lugano, 1991. Also in Computational
Linguistics, Vol. 19, No. 4, 571–590, 1994. 71
[MMdM99] I. Maks, W. Martin, and H. de Meerseman. RBN Manual, 1999.
61
[MMH03] D.J. McArthur, J.R. McDonald, and J. Hossack. A Multi-Agent
Approach to Power System Disturbance Diagnosis. SpringerVerlag,
Berlin, 2003. 167
[Mon74] R. Montague. Formal Philosophy. Yale University Press, New
Haven, 1974. 4
[Moo00a] R. C. Moore. Improved Left-Corner Chart Parsing for Large
Context-Free Grammars. In Proceedings of the 6th IWPT, pages
171–182, Trento, Italy, 2000. 70, 106
[Moo00b] R. C. Moore. Time as a Measure of Parsing Efficiency. In Proceedings
of Efficiency in Large-Scale Parsing Systems Workshop,
COLING’2000, pages 23–28, Saarbrucken: Universitaet des Saarlandes,
2000. 106
[Nev05] Z. Nevěřilová. Visual Browser. http://nlp.fi.muni.cz/
projects/visualbrowser/, 2005. 56, 67
[NM04] M. Neteler and H. Mitasova. Open Source GIS: A GRASS GIS
Approach. Kluwer Academic Pub, 2nd
edition, 2004. 42
[NNH06] Jens Nilsson, Joakim Nivre, and Johan Hall. Graph Transformations
in Data-Driven Dependency Parsing,. In Proceedings of the
21st Conference on Computational Linguistics and 44th Annual
Meeting of the ACL, pages 257–264, Sydney, 2006. 108
[NP01] I. Niles and A. Pease. Towards a standard upper ontology. pages
2–9. ACM Press New York, NY, USA, 2001. 62
[NS04] T. Nagata and H. Sasaki. A Multi-Agent Approach to Power System
Restoration. 2004. 167
[NT04] I. Niles and A. Terry. The MILO: A general-purpose, mid-level
ontology. 2004. 62
221
BIBLIOGRAPHY BIBLIOGRAPHY
[O+
02] Ian Oeschger et al. Creating Applications with Mozilla. O’Reilly
and Associates, Inc., Sebastopol, California, 2002. 41
[OSI84] International Organization for Standardization. Basic Reference
Model for Open System Interconnection, 1984. ISO 8072. 173
[P+
02] Jan Petr et al. Slovník spisovného jazyka českého (Dictionary of
Written Czech, SSJČ). Academia, Praha, 1st
edition, 2002. electronic
version, created in the Institute of Czech Language, Czech
Academy of Sciences Prague in cooperation with Faculty of Informatics,
Masaryk University Brno. 56, 61
[PH05a] M. Prýmek and A. Horák. Design and Implementation of MultiAgent
System for Analysis of Electrical Power Networks. In Proceedings
of ElNet 2005 Workshop, Ostrava, 2005. VSB Technical
University of Ostrava. 162, 180
[PH05b] M. Prýmek and A. Horák. Multi-Agent Framework for Power
Systems Simulation and Monitoring. In Proceedings of ICICT
2005, pages 525–538, Cairo, Egypt, 2005. Information Technology
Institute, Giza. 162
[PH06a] Karel Pala and Aleš Horák. From WEB Pages to Dictionary:
a Languge-independent Dictionary Writing System. In Proceedings
of the 12th
EURALEX International Congress, Turin, Italy, 2006.
10
[PH06b] Miroslav Prýmek and Aleš Horák. New Features in Power Networks
Modelling Using the Rice System. In Proceedings of ElNet
2006 Workshop, Ostrava, 2006. VSB Technical University of Ostrava.
162
[PHL+
06] J. Pustejovsky, C. Havasi, J. Littman, A. Rumshisky, and M. Verhagen.
Towards a Generative Lexical Resource: The Brandeis
Semantic Ontology. In Proceedings of LREC 2006, Genoa, Italy,
2006. demo. 59
[PJ94] K. Pala and Všianský J. Slovník českých synonym (Dictionary of
Czech Synonyms, SČS). Lidove Noviny Publishers, Praha, 1994.
56
[PP02] Tomáš Pavelek and Karel Pala. VisDic – A New Tool for WordNet
Editing. In Proceedings of the First International Global WordNet
Conference, Mysore, India, 2002. Central Institute of Indian
Languages. 25
[PRS97] K. Pala, P. Rychlý, and P. Smrž. DESAM — annotated corpus for
Czech. In Proceedings of SOFSEM’97, pages 523–530. Springer-
222
BIBLIOGRAPHY BIBLIOGRAPHY
Verlag, 1997. Lecture Notes in Computer Science 1338. 7, 70, 86,
92
[Prý04] M. Prýmek. Vybrané problémy návrhu simulátor˚u umělého života
a jejich řešení v systému Kairos (Selected Problems in Artificial
Life Simulator Design and Their Solution in the Kairos System).
In Kognika a umělý život (Cognica and Artificial Life) IV. Silesian
University, Opava, Czech Republic, 2004. 164
[PSH01] D. O. Koval P. S. Hale, R. G. Arno. Analysis techniques for electrical
and mechanical power systems. In Proceedings of 2001 IEEE
I&CPS Tech. Conf., pages 61–65, 2001. 179
[PŠ97] Karel Pala and Pavel Ševeček. Valence českých sloves (Valencies
of Czech Verbs). In Proceedings of Works of Philosophical Faculty
at the University of Brno, pages 41–54, Brno, 1997. Masaryk
University. In Czech. 10
[Rad93] Andrew Radford. Minimalist Syntax. Cambridge University Press,
Chicago, 1993. 69
[Ray07] T. Ray. The Tierra Simulator, 2007. http://www.his.atr.jp/
~ray/tierra/. 164
[Reh03] C. Rehtanz. Autonomous Systems and Intelligent Agents in Power
System Control and Operation. Springer-Verlag, Berlin, 2003. 160,
165
[RK07] Albena Rangelova and Jan Králík. Wider Framework of the Research
Plan Creation of a Lexical Database of the Czech Language
of the Beginning of the 21st Century. In Proceedings of
the Computer Treatment of Slavic and East European Languages
2007, pages 209–217, Bratislava, Slovakia, 2007. 60
[RLN93] P.S. Rosenbloom, J.E. Laird, and A. Newell. The Soar Papers:
Readings on Integrated Intelligence. MIT Press, Cambridge, MA,
1993. 142
[RM98] Louis Rosenfeld and Peter Morville. Information Architecture for
the World Wide Web. O’Reilly and Associates, Inc., Sebastopol,
California, 1998. 40
[RN03] S. Russel and P. Norvig. Artificial Intelligence, A Modern Approach.
Pearson Education, Inc., Upper Saddle River, New Jersey,
2nd edition, 2003. 160
[RUB07] The Ruby Programming Language, 2007. http://www.ruby-lang.
org/en/. 42
223
BIBLIOGRAPHY BIBLIOGRAPHY
[Š04] Pavel Šmerk. Unsupervised Learning of Rules for Morphological
Disambiguation. In Lecture Notes in Artificial Intelligence 3206,
Proceedings of Text, Speech and Dialogue 2004, pages 211–216,
Berlin, 2004. Springer-Verlag. 182
[Sam00] G. Sampson. A Proposal for Improving the Measurement of Parse
Accuracy. International Journal of Corpus Linguistics, 5(01):53–
68, 2000. 115
[Sam07] Geoffrey Sampson. Leaf-ancestor assessment software, 2007.
http://www.grsampson.net/Resources.html. 115
[San04] Yohanes Santoso. Gnome’s Guide to WEBrick, 2004. (http://
microjet.ath.cx/WebWiki/WEBrick.html). 46
[SB03] G. Sampson and A. Babarczy. A test of the leaf-ancestor metric
for parse accuracy. Natural Language Engineering, 9(04):365–380,
2003. 115
[SB95] SC Sciacca and WR Block. Advanced SCADA concepts. Computer
Applications in Power, IEEE, 8(1):23–28, 1995. 161
[Sea69] J.R. Searle. Speech Acts. Cambridge University Press, Cambridge,
1969. 176
[Sed05] Radek Sedláček. Morphemic Analyser for Czech. PhD thesis,
Masaryk University, Brno, Czech Republic, 2005. 71, 82, 109,
182
[SH98] P. Smrž and A. Horák. Determining Type of TIL Construction
with Verb Valency Analyser. In Proceedings of SOFSEM’98, pages
429–436, Berlin, 1998. Springer-Verlag. 10
[SHKC05] Lukáš Svoboda, Aleš Horák, Vladimír Kadlec, and Pavel Cenek.
Language Resources for Intelligent Processing of Dialogues about
Electrical Networks. In Proceedings of ElNet 2005 Workshop, Ostrava,
2005. VSB Technical University of Ostrava. 162
[Sch07] Kim Schulz. Hacking Vim: A Cookbook to get the Most out of the
Latest Vim Editor. PACKT Publishing, 2007. www.vim.org. 19
[Ski57] B. F. Skinner. Verbal behaviour. New York, 1957. 163
[SL00] Rohini Srihari and Wei Li. A Question Answering System Supported
by Information Extraction. In Proceedings of the 1st Meeting
of the North American Chapter of the Association for Computational
Linguistics, pages 166–172, Seattle, U.S.A, 2000. 69
[SLŽ02] M. Straňáková-Lopatková and Z. Žabokrtský. Valency Dictionary
of Czech Verbs: Complex Tectogrammatical Annotation. In C. Paz
224
BIBLIOGRAPHY BIBLIOGRAPHY
Suárez Araujo M. González Rodríguez, editor, LREC2002, Proceedings,
volume III, pages 949–956. ELRA, 2002. 11
[Smi90] B. Smith. Towards a History of Speech Act Theory. de Gruyter,
Berlin/New York, 1990. 176
[SOA03] SOAP 1.2 Part 1: Messaging Framework, June 24 2003. http://
www.w3.org/TR/soap12-part1/. 160
[SP04] Lukáš Svoboda and Lubomír Popelínský. Communication with
WWW in Czech. Cybernetica, 40(3):349–363, 2004. 185
[SPHS05] P. Smrž, M. Prýmek, A. Horák, and A. Sinopalnikova. Emergent
Systems and Intelligent Agents for Simulation of Power Systems,.
In Proceedings of ELNET 2004 Workshop, Ostrava, 2005. VSB
Technical University of Ostrava. 162
[Svo04] L. Svoboda. Multiword expression processing (in Czech). In
Vojtěch Svátek, editor, Poster Proceedings of the Znalosti 2004
Workshop, 2004. 186
[TH01] D. Thomas and A. Hunt. Programming Ruby. Addison-Wesley
Reading, MA, 2001. 42
[Tic04] P. Tichý. Collected Papers in Logic and Philosophy. Prague:
Filosofia, Czech Academy of Sciences, and Dunedin: University
of Otago Press, 2004. 4, 121
[Tic80] P. Tichý. The Semantics of Episodic Verbs. Theoretical Linguistics,
7:264–296, 1980. 136, 140
[Tic82] P. Tichý. Foundations of Partial Type Theory. Reports on Mathematical
Logic, 14:52–72, 1982. 122, 155, 210
[Tic88] P. Tichý. The Foundations of Frege’s Logic. de Gruyter, Berlin,
New York, 1988. 4, 119, 120, 121
[Tic94a] P. Tichý. Cracking the Natural Language Code. From the Logical
Point of View, III(2):6–19, 1994. 209
[Tic94b] P. Tichý. The Analysis of Natural Language. From the Logical
Point of View, III, 2:42–80, 1994. 209
[TK98] J.C. Trueswell and A.E. Kim. How to prune a garden-path by
nipping it in the bud: Fast-priming of verb argument structures.
Journal of Memory and Language, (39):102–123, 1998. 90
[TMBM06] Dan Tufi¸s, Verginica Barbu Mititelu, Luigi Bozianu, and C˘at˘alin
Mih˘ail˘a. Romanian WordNet: New Developments and Applications.
In GWC 2006, Jeju Island, Korea, 2006. Masaryk University,
Brno. 207
225
BIBLIOGRAPHY BIBLIOGRAPHY
[Tom86] M. Tomita. Efficient Parsing for Natural Languages: A Fast Algorithm
for Practical Systems. Kluwer Academic Publishers, Boston,
MA, 1986. 83
[Too07] Field Linguist’s Toolbox, 2007. http://www.sil.org/computing/
toolbox/. 9
[VB+
98] P. Vossen, L. Bloksma, et al. The EuroWordNet Base Concepts
and Top Ontology. Technical Report Deliverable D017, D034,
D036, WP5 EuroWordNet, LE2-4003, University of Amsterdam,
1998. 15, 129, 135
[vdBB02] A. van den Bosch and S. Buchholz. Shallow Parsing on the Basis
of Words Only: A Case Study. In ACL 2002, pages 433–440, 2002.
76
[vdV03] Eric van der Vlist. RELAX NG. O’Reilly Media, 2003. 45
[Vos98] Piek Vossen, editor. EuroWordNet: A Multilingual Database with
Lexical Semantic Networks for European Languages. Kluwer Academic
Publishers, Dordrecht, 1998. 8, 61
[Vyk05] Radek Vykydal. Nástroje pro vývoj gramatik přirozeného jazyka
(Tools for Developing Natural Language Grammars). Master’s
thesis, Faculty of Informatics, Masaryk University, Brno, Czech
Republic, 2005. In Czech. 73
[Vyk07] Radek Vykydal. The Grammar Development Workbench project,
2007. http://nlp.fi.muni.cz/projects/grammar_workbench/
manual-en/. 76
[WN07] WordNet – a lexical database for the English language. http://
wordnet.princeton.edu/, 2007. 8, 135
[Zig05] The ZigBee Alliance, 2005. http://www.zigbee.org/. 210
[Žab05] Zdeněk Žabokrtský. Valency Lexicon of Czech Verbs. PhD thesis,
Faculty of Mathematics and Physics, Charles University in Prague,
2005. 20, 134
[ŽL04] Zdeněk Žabokrtský and Markéta Lopatková. Valency Frames of
Czech Verbs in VALLEX 1.0. In Adam Meyers, editor, HLTNAACL
2004 Workshop: Frontiers in Corpus Annotation, pages
70–77, May 2–7, 2004 2004. 7, 14
226
Annotation
Computer Processing of Czech Syntax and
Semantics
In this text, we present results achieved in several natural language
processing research projects of the NLP Centre, Faculty of Informatics,
Masaryk University. The central topics of all these projects are the
syntactic and semantic analysis of natural language (NL) sentences with
the concentration on the Czech language. The leading person in these
projects is Aleš Horák.
After an introduction, the second chapter describes three years of development
of VerbaLex, a large lexicon of Czech verb valencies in the
form of complex valency frames. This part is then followed by detailed
descriptions of developed tools for working with this as well as other language
resources. The presented tools are VisDic, DEBVisDic, DEBDict,
PRALED and others. These tools are used by project teams all over the
world.
The next chapter presents the latest development of the syntactic
analyser synt that has been under development in the NLP Centre for
several years. Besides the comprehensive description of synt inside and
formats used, we also provide a comparison with several other natural
language parsers, in which we show that the synt qualities are at least
comparable to the best current parsers.
The fourth chapter outlines the advances made in the Normal Translation
Algorithm (NTA) for Transparent Intensional Logic (TIL). It describes
the methods and techniques aimed at an automatic translation
from a NL sentence to its meaning expressed as a construction in TIL.
227
ANNOTATION ANNOTATION
The description is not complete, yet, but we have concentrated on selected
phenomena where we offer sample solutions or even prototypical
implementations.
The last chapter gives details of a project that concentrates on intelligent
methods for increasing the reliability of electrical networks. One
its part involves the development of a human-machine communication
framework for dialogues about the specific knowledge domain of electrical
power systems (EPS). The task of another project part is the development
of a multi-agent system for representing the EPS processes.
These allow simulating different configurations of an EPS setup with an
automatic computation of the economic aspects of the system failures.
Anotace (Czech Annotation): Počítačové zpracování
české syntaxe a sémantiky
Tato práce prezentuje výsledky dosažené při řešení vybraných výzkumných
projektů zpracování přirozeného jazyka v Centru ZPJ na Fakultě
informatiky Masarykovy univerzity. Hlavním tématem těchto projektů je
syntaktická a sémantická analýza vět přirozeného jazyka se zaměřením
na češtinu. Vedoucím uvedených projektů je Aleš Horák.
Po úvodu je ve druhé kapitole popsána dosavadní tříletá práce na
velkém lexikonu českých slovesných valencí VerbaLex, kde valence jsou
uloženy ve formě tzv. komplexních valenčních rámců. Po této části následuje
detailní popis vyvinutých nástrojů pro práci s tímto i jinými jazykovými
zdroji. Jedná se o nástroje VisDic, DEBVisDic, DEBDict, PRALED
a další. Tyto nástroje byly (a stále jsou) používány v jazykových
výzkumných projektech po celém světě.
Následující kapitola ukazuje poslední vývoj syntaktického analyzátoru
synt, který je jedním z dlouhodobých projektů Centra ZPJ. Kromě detailního
popisu vnitřních technik a formátů systému synt uvádíme také
srovnání s několika dalšími syntaktickými analyzátory přirozeného jazyka,
kde ukazujeme, že synt je minimálně srovnatelný s nejlepšími současnými
analyzátory.
Ve čtvrté kapitole shrnujeme pokrok ve vývoji Algoritmu normální
translace (NTA) pro transparentní intenzionální logiku (TIL). Popisujeme
zde metody a techniky pro automatický překlad věty přirozeného
jazyka na její význam ve formě konstrukce transparentní intenzionální lo-
228
ANNOTATION ANNOTATION
giky. Uvedený popis není ještě kompletní, ale zaměřili jsme se na vybrané
problematické jevy, u kterých uvádíme příklady řešení nebo dokonce prototypové
implementace.
Poslední kapitola tohoto textu poskytuje detaily projektu zaměřeného
na inteligentní metody pro zvýšení spolehlivosti elektrických sítí. Tento
projekt zahrnuje jako jednu ze svých částí vývoj komunikačního rozhraní
člověk-stroj pro dialogy ze specifické znalostní domény elektrorozvodných
systémů (ERS). V další popsané části je vyvíjen multiagentní systém
pro reprezentaci procesů ERS s možností simulace různých situací ERS
s automatickým výpočtem ekonomických dopadů výpadků systému.
229
Book orders should be addressed to:
Librix.eu
Gorkeho 41
CZ-602 00 Brno, Czech Republic
E-mail: sales@librix.eu
Computer Processing of
Czech Syntax and Semantics
Aleš Horák
Published by Librix.eu, Gorkeho 41, 602 00 Brno, Czech Republic
in 2008
First edition, 2008
ISBN 978-80-7399-375-7