Encoding UNL Expressions Some Problems and Proposals Igor

Encoding UNL Expressions: Some Problems and Proposals Igor Boguslavsky UNL Russia bogus@iitp. ru

Generals l UNL: an interlingua or not? l No ambiguity l UNL vs. natural language – “at least as powerful as any NL”? – semantics vs. KB l Correct UNL vs. adequate UNL

Adequacy conditions An adequate UNL expression should: l preserve the meaning of the source text; l be convenient for prospective applications, including deconversion.

Possibility of inverse generation? Necessary but insufficient: l invitation of the president l mod(invitation, president)? – the president invites somebody – somebody invites the president l Russian: He received the shower (= took the shower)

Tentative procedure How to develop a definite and common view on what UNL expressions are adequate? 1. UNL from LCs to UNLC 2. Comments from LCs to UNLC 3. Feedback from UNLC to LCs 4. Update of UNL by LCs

Universal Words l Headwords l Restrictions l Attributes

Headwords l Multi-word UWs l Support verbs

Multi-word UWs They should be avoided, if their meaning is representable as a combination of meanings of words they are composed of: l UW to be avoided: «Ministery of foreign affairs» l UW to be preferred: mod(ministery. @entry, affair. @pl) mod(affair. @pl, foreign)

Why so? If any free word combination can be made a UW, one can never hope that other partners will have matching UWs in their dictionaries.

Appropriate multi-word UWs l Non-compositional phrases: – «look for(agt>thing, obj>thing)» – «look like(aoj>thing, obj>thing)»

A convenient compromise l To account for the fact that a phrase is considered as denoting a single concept, the UNL expression can be enclosed in a scope: mod: 01(ministery. @entry, affair. @pl) mod: 01(affair. @pl, foreign)

Another possibility (Ch. Boitet) l Postulate one UW having the internal structure: «mod(ministery, affair. @pl)_mod(affair. @pl, foreign)»

Restrictions in UW/KB Semantic function (2) Knowledge Base function (3) Argument frame function (1)

(1) Semantic function l Restricting the meaning - needed, in particular, to ensure – disambiguation of the head word – selection of the translation equivalent

(2) KB function l Locating the UW in the KB - needed, in particular, to ensure – choice of a nearest UW in the case the direct equivalent is absent in the UW dictionary (replacement ability) – semantic inference

(3) Agrument frame function l Presenting the argument frame.

Correlation between the semantic and the KB functions Semantically- and KB-oriented restrictions do not necessarily coincide: – semantic restriction: book(icl>thing) – KB restriction: titmouse{(icl>bird)}

How to select semantic restrictions l September{(icl>month>date)} l answer(icl>do) (for cases like answer questions) – answer(icl>be) (for cases like answer expectations) – answer(icl>thing) (for cases like know the answer) l Ru: zhenit’sja – marry(agt>male), vyxodit’ zamuzh – marry(agt>female).

VERY IMPORTANT! l Semantic restrictions should effectively distinguish the meaning we restrict from all other relevant meanings of the same English headword. l They should NOT be equally applicable to more than one meaning. l They should be easily understandable.

Example l operator - all the meanings denote a thing l WRONG (in the sense ‘inadequate’): – operator(icl>thing) l CORRECT – operator(icl>human) – operator(icl>abstract thing)

Relations useful for disambiguation l icl l iof l equ l ant (disappeared from the specs? ) – poor(icl>bad): poor quality – ? ? ? poor(icl>having little money) • “having little money” is a bad UW – poor(ant>rich): poor people

Needs to be emphasized again l UNL News 1: build global knowledge – build(agt>thing, obj>thing) l Does not differentiate between different meanings of the headword: – build a railway (a house): build(agt>thing, obj>concrete thing) – build plans (knowledge): build(agt>thing, obj>abstract thing)

KB function l UW: September l MD: September{(icl>month>date)}

What remains unclear-1: KB semantics. Links between related concepts of different semantic categories are missing. There is no way to express the relationship between “dance” (as a verbal concept) and “dance” (as a nominal concept)), “government” and “governmental”, etc. – dance({icl>do(}agt>person{)}) – dance(icl>action{>event})

What remains unclear-2: l Status of UWs within the restrictions: l propose(agt>thing, gol>thing, obj>thing) – They proposed to the president that a special committee should be set up l «set up» does not fall under «thing» . But where then?

What is an argument-1? l. A is an argument of L --> A is integral to the meaning of L.

What is an argument-2? l. A is semantically obligatory: L cannot be semantically defined without A being mentioned. l A is not always syntactically obligatory: it can remain unmentioned in a sentence.

Example: buy has 4 arguments: a buyer, an object, a seller, the money paid. l All of them are semantically obligatory: “buying” cannot exist without any of them. l None of them is syntactically obligatory: l – I bought a book (the seller and the money are not mentioned). – To buy is more pleasant than to sell (no arguments are mentioned).

Semantic roles vs. predicateargument relations l UNL does not mark predicate-argument relations in a systematic way. l Assumption: arguments can be reliably identified based on their semantic role.

It does not work. Why? l Too many «difficult» cases. Only a part of semantic relations between the words can be reliably interpreted in terms of semantic roles. l Too many mismatches. Assignment of semantic roles cannot be done in a consistent way (especially in the UNL multi-lingual and multi-cultural environment).

The reason: Numerous mismatches in the representation of the same or similar phenomena are rooted in the fundamental impossibility of a consistent interpretation of ALL argument relations in terms of a fixed SMALL number of semantic roles.

Examples l Nothing (obj) prevents the members (ben) from discussing (gol) this problem – why beneficiary (ben)? – why finite state (gol)? l protect nature from pollution (? ) l familiarize students with India (? )

Difference between arguments and non-arguments-1 l Any nominal concept can have a purpose, e. g. – a stone for driving nails l Therefore {pur>uw} is assigned to UW «thing» and is inherited by all UWs lying below.

Difference between arguments and non-arguments-2 l Purpose is NOT an argument of “stone”: a stone has no obligatory conceptual link with the purpose. l Purpose IS an argument of “method”: a method cannot exist without a purpose. – a method for calculating taxes

Another example: borrow – X borrows Y from Z for W = • Z owns Y, • X makes Z to give him Y, • X promises Z to give Y back after W expires – borrow cannot exist without 4 participants: agent, object, source, duration

Difference between arguments and non-arguments-3 l Each action has a certain duration. Therefore {dur>time} is assigned to UW «do» and is inherited by all UWs lying below. l Besides this, borrow has a semantic argument with the role ‘duration’

Argument vs. non-argument (1) John borrowed money for 3 years – Argument W: term of the loan. John promised to return money after 3 years (2) John has been borrowing money for 3 years – Non-argument: the situation ‘John is borrowing money’ lasted for three years (the term of each loan is not specified)

Why important? - For semantic processing l (1) can answer the question on the terms of the loan - (2) cannot do so. l mod(invitation, president) – the president invites somebody (arg. 1) – somebody invites the president (arg. 2) – the invitation has an unspecified connection with the president (non-arg. ) l NB: the specs do not allow to draw this distinction!

Why important? - For deconversion l The arguments and non-arguments are very often encoded differently: – dur: Ru «borrow on 3 years» vs. during 3 years – rsn: afraid of darkness, tremble with fear not: *afraid because of darkness – scn: In [scn] this country the relations between the nations are based on [scn-arg] mutual respect

How to differentiate? The distinction between the arguments and the non-arguments should be drawn both in the UWs and in UNL expressions.

Proposal for UWs l Restrictions corresponding to arguments should be systematically and exhaustingly represented in KB. l They can either be included into the UW, or be inherited from upper concepts.

Proposal for UWs l They should be formally opposed to non -argument restrictions. l One of the ways: capitalization. – thing{(and>thing, …, pur>uw, …)} – method(icl>abstract thing, Pur>uw) – do{and>do, …, dur>period, …)} – borrow(icl>do, …Dur>period) l Another possibility: – dur vs. dur. @A

Proposal for UNL expressions l Mark argument relations (Ch. Boitet): – @A, @B, @C… l relation for (1): – dur. @A(borrow, year) l relation for (2): – dur(borrow, year)

Sample UW dictionary entry l Current UW: responsible(aoj>thing, obj>thing) l It is proposed to introduce a comment: responsible (Aoj>thing, Obj>thing, Gol>*) ; he (aoj) is responsible to me (gol) for his actions (obj) (example) ; IB_Ru, 29/11/02 (author and date)

Comments in MD l Not only for illustrating argument frames, but also for clarifying concepts. l Specs modification is needed.

Attributes l Dictionary of attributes (explanation of the attribute, examples) l Procedure for introduction new attributes should be set up.

Issues concerning KB l Adjectival concepts

Adjectival concepts: mod vs. aoj l Two major classes of adjectives: – (aoj>thing) vs. (mod<thing) l Specs: “For an adjectival concept, (aoj>thing) or (mod<thing) should be attached to the Basic UW. (aoj>thing) is for expressing a predicative concept, whereas (mod<thing) is for expressing restrictive concept”.

We should distinguish between: (a) a syntactic property: whether the adjective is used predicatively (Greeks are wise) or attributively (the wise Greeks); (b) a semantic property: what does the adjective mean when used attributively: – restriction; – qualification. Only (b) should interest us.

Restrictive vs. non-restrictive l Wise Greeks diluted wine with water – restrictive: Those Greeks who were wise diluted wine with water. Silly ones didn’t. – non-restrictive (qualificative): Greeks were wise. They diluted wine with water. l Non-attributive (predicative) adjective does not restrict the noun: – Greeks are wise.

Adjectives l Some adjectives can only be restrictive: – Many dogs have curly hair l Some adjectives can only be nonrestrictive: – Get those damned dogs out of the room! – Dear colleagues! l Most of the adjectives can be both restrictive and non-restrictive.

Non-adjectives l The old people in the street were very tired. – Those who were in the street were tired; others were not tired. – The old people were very tired. They were in the street. l There is no UW to which a restriction can be assigned!

What is more… l In some languages restricive vs. nonrestrictive interpretations of relative clauses are marked by punctuation (English, Spanish).

Restrictive l No commas are allowed: – The old people who came a long way were tired. – Los viejos que habían venido de muy lejos estaban cansados.

Non-restrictive l Commas are needed: – The old people, who came a long way, were tired. – Los viejos, que habían venido de muy lejos, estaban cansados.

Proposal l Renounce from the division of adjectives into (aoj>*) and (mod<*). l In order to account for this difference, introduce two attributes (@restr, @nonrestr) which can be added to any modifier (an adjective, a prepositional phrase, a clause), if the author wishes to mark the restrictive or non-restrictive interpretation.

Why attributes are better than restrictions? l Attributes reflect the point of view of the speaker in the current situation and not the permanent property of the word. l Attributes are optional and may be not assigned, if the author does not wish to specify his point of view.

The strongest argument l The restrictive vs. non-restrictive opposition is relevant not only for adjectives but also for other types of noun modifiers. l These modifiers cannot be assigned restrictions but can easily take an attribute.