An Intuitive Representation of Human Languages for Translation Gábor Prószéky MorphoLogic& Faculty of Information Technology, Pázmány University Kalmár.

An Intuitive An Intuitive Representation of Human Representation of Human

Languages Languages for Translationfor Translation

Gábor PrószékyGábor PrószékyMorphoLogicMorphoLogic

&&

Faculty of Information Faculty of Information Technology,Technology,

Pázmány UniversityPázmány University

Kalmár WorkshopKalmár WorkshopSzeged, October 1-2, 2003Szeged, October 1-2, 2003

Contents

Some words on Prof. Kalmár’s activity in computational linguistics

Problems of human language description with formal tools

A new representation with patterns Introduction to machine translation

methods Application of patterns to

translation

Kalmár Kalmár Workshop Workshop

2003 2003

Gábor Prószéky:An Intuitive

Representationof Human Languages

for Translation

Kalmár & languages

Kalmár’s paper in formal language theory: „An Intuitive Representation of Context-Free Languages”

Kalmár’s activity in machine translation (conference in 1962): „Representation of Languages with the Help of Mathematical Structures”


2003 2003



for Translation

Linguistic representation problems of the 60’s

Dependency structure Constituent structure X-bar theory:

X’ (P) X (Q) Related structures Using transformations


2003 2003



for Translation

Structured symbols

Linguistic categories: atomic symbols

Not enough: subcategorization Semantic features: ± alive, ... Syntactic features: ± countable,

... Rule sets instead of rules ID/LP


2003 2003



for Translation

Feature structures

DAGs Unification problems Feature geometry, typed

features LFG, GPSG, HPSG Parsing: CF-skeleton +

features or feature structures only?


2003 2003



for Translation

Complexity of NL grammars

RG/FSA: not enough CF/RTN: not enough CS ? 0/ATN: Turing Machine Transformations and

metarules Arguments for and against


2003 2003



for Translation

NL grammar formalisms Competence and performance? Kornai number (left-recursion, center-

embedding, “respectively” construction) Gradually from unrestricted to regular (i) anbn ->a*b* (n is lost!) (ii) anbn ->{ε,ab,aabb,aaabbb} “Finitization” by length No structure in FSA; finite systems,

however, can produce structural output


2003 2003



for Translation

Syntax and semantics

Logical representations(e.g. λx.dog(x), λx.run(x))

World-knowledge representations(e.g. IS-A, PART-OF, INSTANCE-OF)

Categorial grammar: early logical representations of syntax (Kalmár)

DCG: interpretation & representation

Rule-to-rule hypothesis


2003 2003



for Translation

Conflict handling

Lexicon meets syntax: who is right?

Lexicon: off-line info coming from past experiences

Which is more important in a specific situation?


2003 2003



for Translation

Open classes

Open vs. closed classes:that is, features can or cannot be overridden

Proper names, jabbers, folk etymology, loanwords, ...

Grammar of closed classes:minimal grammar


2003 2003



for Translation

Finite morphology Finite patterns Finite number of entries Descriptions assigned to

entries Finite & open vs.

infinite & closed Underspecified entries for

guessing



for Translation

Finite syntax

“Item and arrangement” (as in morphology)

“Arrangement” describes a rather free constituent-order

Metawords in a meta-dictionary, e.g. ‘(Det (Adj (N)))’ ‘DAN’

Cascades without loop


2003 2003



for Translation

The „plastic box”

John is a boy. ”John” is a noun. Go is a verb. ”Go” is a verb. is a sign. ”” is a sign. is a . � �

(where is a ”plastic box”)�


2003 2003



for Translation

Real examples

(a) Unusual use:Go is a verb.POS [np] POS [v]

(b) Metaphor:My car drinks a lot.ANIMATE [+] ANIMATE [-]

(c) Unknown entry:Kalmár is a family name.POS [np]


2003 2003



for Translation

Linguistic frames

Psychology: ”Gestalt” Morphological complex

structures treated as frames by humans

Frames in AI: ‘shopping’, ‘walking’, ...

As ‘high-level parsing’ relates to ‘detailed on-line analysis’


2003 2003



for Translation

Translation of human languages

old problems (50’s) direct (60’s) interlingual (70’s) transfer (80’s) examples (90’s)


2003 2003



for Translation


2003 2003



for Translation

Patterns: general linguistic Patterns: general linguistic informationinformation in lexicalized formin lexicalized form Short, fully specified patterns are:

lexical entries Longer, fully specified entries are:

multi-word expressions Partially underspecified patterns are:

collocations, phrasal verbs, idioms Totally underspecified patterns are:

linguistic rules Pattern/interpretation pairs:

Translation Description Language


2003 2003



for Translation

The MetaMorpho principlesThe MetaMorpho principles

No single words but contextual expressions (in form of patterns) only

Pattern pairs: input/interpretation structure pairs

Single pass: no separate transfer steps Target structure generation:

by-product of parsing

Jabberwocky

‘Twas brillig, and the slighty tovesDid gyre and gimble in the wabe:All mimsy were the borogroves,And the mone raths outgrabe.


2003 2003



for Translation

‘Twas �, and the � �sDid � and � in the �:All � were the �s,And the � �s �.


2003 2003



for Translation

Translation rules for Jabberwocky

‘twas � � volt �, and � �, és � the �s did � a �ok �tak � and � � és � in the � a �ban all � teljesen � � were the �s �k voltak az �ok the �s � a �ok �tek


2003 2003



for Translation

‘Twas �, and the � �s

Did � and � in the �:All � were the �s,And the � �s �.

� volt, és a � �ok�tak és �tek a �ben:teljesen � voltak a �okés a � �ok �tek.


2003 2003



for Translation

Translation of Jabberwocky

Dzsebervoki

Brillig volt, és a szlájti tóvokgájertak és gimbeltek a vébben:teljesen mimszik voltak a borogróvokés a món rátok autgrébtek.


2003 2003



for Translation

An intuitive representation...

1. X-bar based structures2. Feature-based descriptions3. Metarules (used off-line) 4. Rule-to-rule principle5. Lexicon should be finite but

open6. Closed classes belong to the

minimal grammar7. Minimal grammar describes

”basically” linguistic elements


2003 2003



for Translation

An intuitive representation...

(cont’d)8. Linguistic constructions can be

described by finite patterns9. A huge & finite description set

is used rather than a limited & infinite grammar

10. In case of conflict, lexical information is either redundant or contradicting to the actual description

11. Known constructions need no real-time analysis (Gestalt, frame)


2003 2003



for Translation

An intuitive representation... (cont’d)

12. ”Broken” frames are analyzed real-time

13. Structural (source/target) pattern pair is assigned to every frame to be translated

14. Target structure is computed while parsing source structure


2003 2003



for Translation


2003 2003



for Translation

An Intuitive Representation of Human Languages for Translation Gábor Prószéky MorphoLogic& Faculty of Information Technology, Pázmány University Kalmár.

Documents

hpsg t

length t

regular t

t finitization

t cfrtn

construction t

contents t

s t examples