Toward A Multiple Environments Model Of Natural Language

TOWARD A MIJI!l'IPL& ENVIROitCI:Nl'S MIJilEL OF lWl'URAL .l.Al«llWl.

Janusz Stanislaw Bie~ Institute of Matheatical Machinea1 University of Warsaw Warazawa1 Poland

Abstract

The idea or treating utterances as programs to be run in hWIBn brains or on a computer is pursued by advocating the use of the environment notion tor natural language semantics. One structure or environments is devoted to keeping track or the real or fictive interlocutors and another to dis­ tinguishing somebody's linguistic behavior from ita pretense and belief'1 while the third one represents the structure or topics in a discourse, The reguired flexibility or environment 118nipu­ lation 1e expected to be supplied by the Bobraw­ Wegbreit primitives.

I.ntroduction

It is now obvious that the huan ability to use language is related closely to intelligence it­ self. Nevertheless1 the complexity or natural language is still rather underestimated by lin­ guiatics1 including in some measure also computa­ tional linguistics. That underestimation results usually in using relatively primitive tools for a !oral description or language. Winograd [ 14] claims rightly that the best teat or a complex model of natural langua&e is to implement it as a langwlge understanding system. There is also a IDOre radical approach1ineisting that natural lan­ guage texts are just programs to be run in our brains [ 9] · Some interesting analogies between language understanding and running a program have been shown1 e. g, in [4], Considering all utter­ ances as imperatives is not entirely a new idea; it can be found also in linguistics papero, e. g. in [13]. But if we treat the utterances as prog­ rams1 we can describe their semantics by means or notions nuch IDOre soJiliaticated and exact than were ever used in linguistics. The purpose of the paper is to advocate the use or the llllltiple en­ vironments model or Bobrow and Wegbreit [2] tor natural language semantics. The basic ideas of this approach will be shown by discussing some well-known problems or reference and presuP­ position. In the early stage of this in:luiry s develoiJDent1 the paper tends to a rather the­ oretical bias, The Bobrow-Wegbreit primitives do not intervene directly into its content, but the flexibility or environment manipulation supplied by them is a necessary condition tor the f'easi­ bili ty of the present approach, which is expected to give valuable guidelines for designing dis­ course understanding systems1 including literary discourse like children's stories treated by

Charniak [3].

·Notion of discourse

The notion or discourse is a rather vague one; therefore1 it is necessary to state What I mean

by the term. I use it as a synonym tor a coherent text and I recognize tour possible aspects or coherency1 which correspond intuitively to dif­ ferent types or acceptability.

First1 there is contextual coherence, realized in a text by inter-sentence and inter-phrase links represented by ·specialized linguistic means which I call pointers. They can be diatinguished1 IDOre or less arbitrar1ly1 in every language, They entail first or all the pronouns (in a large aen­ se1 including pro-adverbs etc.). The second iJD­ portant type or pointers consists of preauppoai­ ti one 1 which di rrer from the pronouns in two ways : they rarely occur on the surface of the text, and they point exclusively to propositions. There are also pointers peculiar to given langu1143ea; e, g., for English it is reasonable to treat the Past Tense as a pointer, following the approach of McCawley [10] and Isard [7]. Now, if there is a pending pointer (without a value) in a text1 the text is not contextually coherent. There is also situational coherence. i. e, the coherence or a text with the situation or its particular usage; e, g.,

1. Be caref'ul1 he might bite you,
2. is perfectly acceptable it the addressee is near a dangerous aniDBl. Introducing a narrator can change such a text into an contextually coherent one1 by substituting its indices (referring to the elements or the real world) back into point­ ers (referring to IDOre or leas abstract text elements), When the addressee has to apply his knowledge or the real world to interpret the text as a coherent one1 I use the notion of relative coherence. This is the case in the text
3. Ann's eldest son has lett Warsaw tor a scholarship study in the Sorbonne,
4. France is an interesting country to study in.

where the knowledge that the Sorbonne is a French university supplies the missing link. In general we can call a text coherent relative to a given domain or knowledge.

Most texts which are incoherent according to the already mentioned rules can still be regarded as coherent ones if the addressee wishes to do so. I call such a ·situation volitiorary coherence. The addressee· s will can be expressed in tvo ways, First, he can supply himself the values for pend­ ing pointers. In particular1 the events required by the presuppositions are freQ.uently accepted as true, as has been pointed out by e, g. Beller.t [1], Wierzbicka [13], recently Karttunen [8], and illustrated by examples:

1. I would like to introduce you to my wife.

• ( 5 ) We regreat that children cannot accoaapaey their parents to commencement exercises.

Vhere the presuppositions carry brand new infor­ mation. SecoDUy, the addressee can decide to consider pending pointers and iaUcea as not relevant to the matter; this procedure seeiiiS to be applied only far some literary texts.

We will discuss nov some problema of reference. Let us start troll the assumption that every per­ son (hUDBn being or robot) vho uses natural lan­ guage perceives also some aspects of the current situation,in particular he recognizes the persons engaged in the discourse and usually realizes possible differences between his own perception of the situation and the perceptions of others. Restating this in llllltiple enviro~Dents teriiiB, ve assume that a special type of enviro~Dents is used to represent the varld as perceived and known by a person;a global environment represents the person under consideration and local envi­ ronments entailed in it are i-.gea of ita inter­

locutors.

An utterance is received together with the infOriiiBtion about its sender and ita addressee; therefore, it can be run in such a vay that the personal pronouns ·r-, ·ve·, ·you· etc. trans­ lated into variables are evaluated respectively in the sender and the addressee environments, yielding their identification. In this vay the evaluation of the sentence ( 6) said by Fred to Stanley:

1. I like your book.

supplies the value eQ.uivalent to the utterance

1. Fred likes Stanley's book.

run in the proper environment. Which environ~~ent is proper depends on the robot's attitude to the discourse. If he believes Fred and assumes Stanley's knoWledge to be basically identical with his ovn, he may run (7) in the environ~~ent of himaelt. a.J.t if he vanta to keep track ot all the possibilities, he has to allocate in Fred's environment Vhich holds mainly the data about his behavior 1 a nev environment far Fred aa described bf Fred and evaluate (7) in it to obtain Fred s Jl'!teMe, and then to evaluate (7) again in Stanley# s knowledge environment to obtain ita probable effect on Stanley's m1nd. The merit ot the approach lies in the tact that it can be applied recursively in the case of direct and indireet speech. z.. g., in the text

• ( 8 ) Frank said:
• ( 9) "When I came back, John vas already waiting far IllS and asked:
• 'How are you~·.
• ( 8) should be run both in the current addressee aai sender environments,(9) in the enviro~nt ot the sender and the new environment ot the Frank as quoted by the sender , and ( 10) in the Frank as quoted by the sender and John as quoted by

Frank as quoted by the sealer. The interpretation ot utterances with ia:Urect speech, e. g.,

1. Fred said to Stanley he liked his book.

and in general the utteraces with the 'rd person pronouns, is 1111ch IIIQJ'e complicated. To account for it ve will introduce auxil1ery mraae emi­ rogeents for running some utterances, with a special variable s1m1lar to the context ot COIOO:VJIR [ 11 ] · The variable Jllllled TOPIC haa a tuple as a value and vhen accessed yields the first element of it, but vhen accessed because of backtracking yields conaecuti ve elements of the tuple, and after exhausting it takes the tuple of TOPIC trom ita super-environment. During a run of an utterance, TCPIC is treQ.uently IIIOditied by adding and per1a1ting its elements. The ele­ ments of TOPIC are pointers to enviroDIIIenta ot (real or fictive) persona and other data struc­ tures representing objects, and also pointers to other }ilrase environment with his own TOPICs · The interpretation ot a text with indirect speech and also single sentences with the 'rd person pronouns can be described now in the tollowiDS way.

1. Every access to a TOPIC element resets TOPIC in such a vay that the accessed element becomes the first one in the tuple.
2. Pronouns in a ain clause and proper DBIIIeS execute immediately the access to the TOPiC variable.
3. Execution of the access reQ.uired by pronouns in a subordinate clause can be auapeaied until the end ot the ain clause processing, it there is no sui table value tor them in the local TOPIC. D. The verbs · say· 1 "describe· etc. create a clause environment with his ov.nTOPIC variable and put the pointer to it as the first element ot TOPIC in the environment of the verb.
4. The subordinate clauses introduced by · say" 1 ·describe" etc. are evaluated in the envirolllllent taken trom the TOPIC of the Jain clause envi­ ronment.
5. Let us see with examples how the rules given above work.
6. ( 12) Attar John lett his apartelll'lt, he vent to the pool hall.

John sets TOPIC to the pointer to the John environment; his is evaluated because there is already sui table value in TOPIC; he eva­ luates also to John.

• ( n) After he lett his apartment, John went to the pool hall.

.After introduces the subordinate clause, which is being evaluated, but it is suspended before executing access by the proDOUnB, because ot the lack ot sui table values tor them in TOPIC · John seta TOPIC, after the end of the main clause processing the subordinate clause is resumecl and the pronouns are properly bound to John.

• ( 14) John went to the pool room after he lett his apartment.
• No problem; John sets TOPIC and the pronouns

• ( 15) He went to the pool room after John lett his apartment.
• He is evaluted ianediately because it is contained in the main clause; therefore it is already bourn when John sets TOPIC. The strong feeling that the value of he should be differ­ ent from John can be explained in terms of computational efficiency: there is no reason to set TOPIC to John if it 1s already the value.
• ( 16 ) What did John say about Dick~
• He said that
• ( 18) he looked like a drunken giraffe on ice slaltes.

\n (16) TOPIC is set first to John, then ·say creates a phrase environment with its TOPIC ini­ tialised to Dick, and puts the pointer to it at the top of its own TOPIC. In ( 17) · he"' evaluates to John, then · safd" recovers from TOPIC the clause environment and evaluates ( 18) in it. Therefore the "he" of (18) is bourn to Dick.

I9def'inite and definite descriptions

Indefinite descriptions of' the type

1. It is a foreigner Who is delivering a speech now.

have clear meaning in our model. They are just declarations, used to create appropriate envi­ ronment (in general, data structures) initialized to the values supplied by the description; the pointer to the new environment is put into TOPIC.

Therefore in the text

1. One young boy has flunked his matriculation exam.
2. He decided to give up further studies.

the "hi of' (21) can receive the proper value of' the young boy mentioned in (20 ). Such an interpretation of' indefinite descriptions has in fact all properties of' the Ref operator intro­ duced f'or the purpose by Bellert [ 1 ); it agrees also with intuitions of' linguists, e. g. Wierz­ bicka [ 13). Def'ini te descriptions are interpreted as functions which When evaluated in a proper environment yield the required value. The inter­ pretation supports the intuition of' Donnellan [ 5) Who claims rightly that a definite descriptions does not in itself' refer to anything but only its use points to a referent. That interpretation explains also why we can utter

1. Her husband is kind to her.

am be understood even if the man is not in fact the husband of the woman. The functions representing definite descriptions can be also just memorized; such a memorizing corresponds to the attributive use (the term of Donnellan) of' the definite description, as, e. g., in~

1. The Smith murderer is insane.
2. in the sense of'
3. Whoever has IIIU"dered Smith, he is insane.
4. To summarize our discusion of reference, let us run a sentence
5. A boy who saw her kissed a girl Who knew him.

First, an environment for "boy· is created and the pointer to it is put into TOPIC. Next, "Who saw her· is evaluated except the "her" because of the lack of' a proper value for it in TOPICJthere­ fore, the clause is suspended. Then the rest of' the main clause is evaluated am an environment for · girl" is lllllde, and the TOPIC is respectively changed. Now the suspended clause can be resumed and eval\\ated in parallel with the clause · who knew him both the pronouns will be properly bourn.

Presuppositions

Presupposi tiona are treated as patterns of' event representations; depending on the volitionary attitude to the text 1 the pattern can be matched against the memory to retrieve an existing event or the appriopriate event can be added to the memory. We express the difference between factual

and non-factual sentences like

1. John regrets that it is raining.
2. John thinks that it is raining.

assuming that the presupposition "it is raining· is evaluated in the environment of' the narrator in the case of' (26) and in the environment of John's belief' in the case of (27). It may seem strange to talk about presuppositions of non­ factual sentences but it is reasonableJ although we do not expect such a presuppositions to be true, we still expect it to be coherent, as in (27), with somebody 1 s belief'. The environment of somebody's belief' and the mentioned earlier environment of' pretense interact with themselves, e. g. in.

1. Fred is lying When he says he likes Stanley's book.

.. book in Fred s fll"etense and to Fred does not like Stanley's book in Fred's belief'.

Which evaluates both to "' Fred likes Stanley's . .

The idea of running an utterance proves use­ ful also in the case of presupposi tiona. The static treatment by projection rules of the relation between the presupposition of' a complex sentence as a Whole and the presuppositions of' its constituent clauses has not been successful. The recent, more dYnamdc, approach of' Karttunen [8) can be easily integrated into our model. E. g., in the following sentences

1. If' Dean told the truth, Nixon is guilty too.
2. If Haldeman is guilty, Nixon is guilty too.
3. If Miss Woods destroyed the missing tapes, Nixon is guilty too.
4. the antecedent clauses should be run before the

consequent clauae to obtain right reaul.ts. In C~o ) the presuppolli tion ot the consequent clause 1a satisfied by the antecedent and in ('1) it can also be the caae,dependins on tbe other know­ ledge present in the enviroraent ot the evalu­ ation.

The JZ'esented approach requires much further development to become a formal a::ldel. Never­ theless, the author feels strons!y that the ideaa ue worth pursuins. Their appeal. lies in the fiexibUity ot possible tr·t·nt ot a sen­ tence, which ·Y account tor different amounts ot attention JUt to it Wlderstandins, differences in heuers attitude, knowledge etc. It is hoped that the nexibil1 ty can be also exploited in other way, i. e., to siaplify the a::ldel as fill' as possible to obtain an efficient implementation.