Natural Language Processing (PA153)
Pavel Rychlý, Karel Pala
Natural Language Processing at FI
Natural Language Processing Centre
around 10 PhD students
you can be part of it
Pavel Rychlý
head of NLP Centre
corpora, lexicography, machine translation
Technical information
Exam: written – 10 questions
open books
max 50 points
25 point to pass
extra points (max 25) for homeworks, projects
find good examples, illustrations to improve understanding
code, language, pictures
Previous knowledge
no special requirements
reading mathematics
probabilities
examples in Python
NumPy, PyTorch (matrix operations)
complements IB030, IB047
Natural language (NL)
Czech, English
not formal languages (programming)
1000s different languages, sub-languages
text
speech
Motivation
Why to pay attention to natural language?
Language behaviour represents one of the fundamental aspects
of human behaviour,
NL is an essential component of our life as a main tool of
communication,
In NL we express and record our knowledge, scientific findings,
world understanding,
NL is a starting point for artificial (formal) languages
Language texts serve as a memory of mankind for knowledge
transfer between generations
NL is a base for human-computer communication
Terminological remark
Used terms:
Quantitative and statistical linguistics
Algebraic linguistics (N. Chomsky)
Mathematical linguistics
computational (počítačová, komputační) linguistics
Today Natural Language processing (ZPJ, NLP)
Human language technology (HLT)
speech processing (ASR, TTS)
What NLP includes?
NL study and research is interdisciplinary:
In linguistics (structural, mathematical)
In psychology and psycholinguistics
In philosophy and logic – relations to the universe of discourse,
reasoning (inference), basic units are truth functions (výroky)
and propositions
In algebraic (later computational) linguistics (in sixties) key
role was played by N. Chomsky (Synt. Struct.)
Language theory in the form of algorithms, and data structures,
large empirical data (corpora)
Relations to the Artificial Intelligence and Cognitive science
Computer instruments for NL – language engineering
NLP – relation to computers
Need for a two-way communication human-computer
So far H-C communication is mainly one-way
A richer H-C communication interface is necessary
NL interface should be smarter and more flexible, especially for
common users
Distinct commercial consequences for the computer market
Influence to the shape of operation systems
Our knowledge about NL structure is incomplete
Relevant role is played by the relation of theory (research) and
applications
NLP – applications 1
Machine translation – testbed for NLP theory
Georgetown–IBM experiment (1954) – demonstration
ALPAC report (1966)
Google Translator – first widely used
Hard task but human quality in some areas
NLP – applications 2
Text processing – spell checkers, grammar and style checkers
Hyphenation
Fulltext search (lemmating, stemming)
Semantic web – intelligent searching, exploiting metadata
Morphological and syntactic analyzers, semantics
Machine readable dictionaries, ontologies
Information extraction
summarization
NLP – applications 3 (speech)
Speech communication with computers (robots)
Synthesis – Text to speech systems (Demosthenes)
Automatic speech recogition (ASR), dictating machines, smart
phones
Applications at courts, in Parliament, in medicine
Can we have a chat with our computer? See PEPPER!
NLP – applications 4 (relation to AI)
Expert systems – e.g. Mycin (diagnostics in medicine)
Dialogue and question-answering (QA) systems
Turing test (Eliza, Loebner Prize, November 2019)
NL understanding in general, stories and messages
Robotic applications – SHRDLU, 1971 (T. Winograd), the first
system containing knowledge, inference and grammar,
Robotic family NAO, PEPPER, ROMEO (Softbank, demo)
Ontology and concept systems for particular domains, sémantic
networks (WordNet)
Problems with NLP
Zipf‘s law
high number of low frequent items (words, phrases, . . . )
Ambiguity
meaning depends on context
Variability
languages evolve
new words/phrases
transfer from other areas
Approaches to NLP
symbolic
rules from experts
no data
statistical
structure/model from experts
optimization of parameters from data
some data
neural (deep learning)
everything from data
huge amount of data
usually a combination
Levels of language analysis
Phonetics and phonology, speech signal
Morphology – flection (and word formation)
Syntax – constituent, dependency
Sémantics – lexical, logical
Pragmatics – relations of users to the language expressions
Discourse, anaphorical relations, reference
Levels – phonetics, phonology
sounds of language (phones)
physical properties of the speech signal
Phonology – phonemes – abstractions on sounds
The smallest units distinguishing meaning, pas – pás
Phonological oppositions: long – short: vola/á
Link to the ASR – automatic speech recognition
TTS (text to speech) – speech synthesis, many systems
ASR (dictation systems, (ARŘ, Newton DS 5, 6)
Intensive research, IBM, Nuance, a lot of money in it
Summary
Problems with NLP
Zipf‘s law
Ambiguity
Variability
Approaches
symbolic (rule-based)
statistical
neural (deep learning)