Hierarchical Graph for Machine Translation
Vít Baisa
DTEDI seminar
spring 2011
Vít Baisa (DTEDI seminar) Hierarchical Graph for Machine Translation spring 2011 1 / 9
Motivation
Language resources for MT questionable (WordNet, VerbaLex, PDT).
Context is crucial but we cannot handle it properly.
Much data + simple algorithms vs. sparse data + complex algorithms.
Even fundamental concepts are vague: what is word, meaning,
(well-formed) sentence, good translation?
People can talk and even translate without any linguistic knowledge.
Everything we know we have learned.
Meaning is not intrinsic: structure of a sentence does not suffice:
Žvoulal si krákornul mášnou kulínou.
Vít Baisa (DTEDI seminar) Hierarchical Graph for Machine Translation spring 2011 2 / 9
State of the Art general purpose MT
MT systems:
Rule-based (morphology, syntax, semantics),
statistical (Google),
hybrid (Yahoo!),
neural networks, . . .
Vít Baisa (DTEDI seminar) Hierarchical Graph for Machine Translation spring 2011 3 / 9
Known issues of MT
Tokenisation: doesn’t → does + n’t or doesn + ’t?
Lemmatisation: neměl → mít or nemít?
Morphologic analysis: ženu → hnát or žena?
Syntactic analysis: Karel mluvil o sexu s Marií.
Word Sense Disambiguation: silný čaj → powerful tea.
Named entities: Včera jsem viděl Královu řeč.
Multiword expressions: vysoká škola → high school.
Metaphors, idioms: Bez práce nejsou koláče → No pain no gain.
Anaphora: Dej ji do vázy. → Put her in a vase.
Vít Baisa (DTEDI seminar) Hierarchical Graph for Machine Translation spring 2011 4 / 9
Language as hierarchy of language units
No distinction between morphology, syntax, semantics and reasoning.
Elementary units are characters or phonems.
Only one relation: (s, t) ∈ R if s is said together with t.
Inductive definition of language:
n → a
má → ma
mám → hlad
odešel → (protože → musel)
když se nenamažeš krémem → spálíš se
Very robust: (3 → (+ → 7)) → (1 → 0)
Equivalent with lambda calculus?
Vít Baisa (DTEDI seminar) Hierarchical Graph for Machine Translation spring 2011 5 / 9
Hierarchical graph
Nodes are language units.
If s and t are LU then s → t is LU.
Several types of edges:
Constituency: mám → hlad
Equivalency: bych would, ps → dog
Partial forward constituency: m hlad
Meaning: meaning(s) = set of neighbours of s in graph.
Synonymy: synonymous(s, t) ⇔ meaning(s) ∩ meaning(t) = ∅
Vít Baisa (DTEDI seminar) Hierarchical Graph for Machine Translation spring 2011 6 / 9
Properties I
Formal approach.
Absolute majority of words (and all sentences) can be divided into
two parts.
Easy knowledge representation: Johann Sebastian Bach se narodil
v roce 1685.
Upper levels are equivalent for various languages (we all think in very
similar ways) – interlingua.
Simple treatment of complex grammatical constraints and
constructions:
Chci (aby to věděl) → I want (him to know)
Nic nemám → I have nothing
Dictionary and grammar within single data structure.
There is linear increase between neighbouring levels.
The more we know the better we memorize (latin, music, math).
Vít Baisa (DTEDI seminar) Hierarchical Graph for Machine Translation spring 2011 7 / 9
Properties II
Passive vs. active knowledge of vocabularies.
Cimrmans theory of externalism.
We understand if we know context: XYZ s r. o.
Synonymy on all levels:
ý – ej,
pěkný – hezký,
Odejdi – Běž pryč.
Meaning on all levels:
í
eji
Karel
nejím maso
Discreteness on all levels:
mš
bát
Petr vyřešil
Vít Baisa (DTEDI seminar) Hierarchical Graph for Machine Translation spring 2011 8 / 9
Goals
Get learning data: we need simple phrases.
Building and tuning of hierarchical graph for Czech and English.
Implement simple algorithms for learning, understanding and
translating.
Standard evaluation + manual evaluation by comparing with
state-of-the-art MT systems.
MT between very different languages (Hungarian, Japanese . . . ).
Derive standard relations and rules from the graph (synonymy,
hyperonymy, subject predicate agreement, . . . ).
Vít Baisa (DTEDI seminar) Hierarchical Graph for Machine Translation spring 2011 9 / 9