A Pragmatics-Based Approach to Understanding

A PRAGMATIC~BASED APPROACH TO UNDERSTANDING INTERS~NTENTIAL ~LIPSI~

Sandra Car berry Department of Computer and Information Science University of Delaware Nevark, Delaware 19715, U3A

ABSTRACT S~A~R-I : "The Korean Jet shot down by the Soviets was a spy plane." IntersententAal eAlipti caA utterances occur frequently in information-seeking dielogues. This FI: "With 269 people on board?"~ paper presents a pragmatics-based framework for F2: "With infrared cameras on board?" interpreting such utterances, ~ncluding identAfi- cation of the spoa~r' s discourse ~oel in employ- Previous research on ellipsis has neglected to ing the fra~ent. We claim that the advantage of address the speaker's discourse Eoals in employing this approach is its reliance upon pragmatic the frasment but reel understanding requires that information, including discourse content and these be identified (Mann, Moore, and Levin, 1977) conversational goals, rather than upon precise (Webber, PoZlack, and Hirschberg, 1982). representations of the preceding utterance alone. In this paper, we investlgate a framework for interpreting Intersententlal ellipsis that occurs INTRODOCTION in task-orlented dialogues. This framework includes: The fraRmentary utterances that are common in communication between humans also occur in man- [1] a context mechanism (Carberry, 1983) that Nachi~e OOmmUlLCcation. Humans perslat in using builds the information-seeker, s underlying abbreviated statements and queries, even in the plan as the dialogue progresses and differen- presence o/ explicit and repeated instructions to tiates be~een local and global contexts. adhere to syntactically and semantically complete sentences (Carbonell, 1983) • Thus a robust [2] a discourse component that controls the natural langua@e interface must handle ellipsis. interpretation of ellipsis based upon discourse goal expectations ~eaned from the We have studied one class of elliptical dial o@ue ; this component "understands" utterances, Intersentential fragments, in the con- ellipsis by identifying the" discourse goal text of an Information-seeklng dialogue. As noted which the speaker is pursuing by employing by Allen(1980), such utterances differ from other the elliptical fragment, and by determining forms of ellipsis in that interpretation often how the frasment should be interpreted rela- depends more heavily upon the speaker's inferred tive to that goal. underlying task-related plan than upon preceding syntactic forms. For example, the fcllowlng [3] an analysis component that suggests possible elliptical fra@ment can only be interpreted within associations o£ an elliptical fragment with the context of the speaker's goal as communicated aspects of the inferred plan for the in the first utterance: information-seeker.

[EX1 ] aT want to cash this check. [4] an evaluation component which, 51yen multiple Smell bills only. * possible associations o£ an elliptical fragment with aspects of the information-seeker,s Furthermore, intersententiel fragments are often underlying plan, selects that ansociation employed to communicate discourse 8oals, such as most appropriate to the discourse context and expressing doubt, which a syntactically complete believed to be intended by the speaker. form of the same utterance may not convey as effectively. In the following alternative responses to the initial statement by SPEAKER-I, INTERPRETATION OF INTERS~qTENTIAL ~LLTPSIS F1 expresses doubt regarding the proposition seated by 3PEAEZB-I whereas F2 merely asks about As Ltlustrated by [EX1], intersententiel the jet's contents. eA1ipticaA fra@ments cannot be fully understood in and of themselves. Therefore a strate8~ for • This work has been partially supported by a interpreting suc~ fra@ments must rely on knowledge grant from the National 3cAence Foundation, XST- obtained frcl sources other than the fragment 8311~00, and a subeontraot from Bolt Beranek and itself. Three possibilities exist: the syntactic Newmm'l Inc. of a grant flwm the Nationa~ ScAence Foundation, T~T-8~19162 ee Ta.~n fr'~ Flowers and Dyer(198~)

188 form ar precedlug utterances, the seaantlo tlon strategy vurke well. In more complex representation of preceding utterances, and expec- domains, it Is necessary to identify the particu- tations gleaned from understanding the preceding lar aspect of the speaker's overall task-related disQourse. plan addressed by the clliptlcal frasment in order to interpret It properly. More recently, Litman The first two strategies are exemplified by and Allen(198q) have extended Allen's model to a the work oC Carbosoll and Hayes(1983), Hendrlx, hierarchy of task-plans and meta-plans. Litman is Sacerdot¢, and Sloc,~( 1976), Waltz( 1978), and currently studying the interpretation of ellipti- Velschedel and 3ondhelmer( 1982 ). Several limita- cal frasments within this enhanced framework. tions exist in these approaches, includiug an ina- bilit 7 to handle utterances that rely upon an In addition to the syntactic, semantic, and assumed communication of the underlying task and plan-based strategies, a few other heuristics have difficulty in resolving ambiguity ="oug multiple been utilized. Carbusoll(1983) uses discourse interpretations. Consider the following two expectation rules that suggest a set of expected dislo~e sequences: user utterances and relate elliptical f~a~ents to these expected patterns. For example, if the sys- SPEAE~R: "I want to take a bus. t~a asks the user whether a particular value The cost?" should be used an the filler o£ a slot in a case frane, the system then expects the user's utter- SPEAKER: "I want to purchase a bus. ance to contain a confirmation or disson~Irmatlon The cost?" pattern, a different filler for the slot, a com- parative pattern such as "too hard", and so forth. Zf a semantic strategy is employed, the case frame Although these rules use expectations about how representation for "bus" may have a "cost of bus" the speaker m~ght respond, they seem to have llt- and a "cost of bus ticket" slot; a~hlgulty arises tle to do with the expected discourse goals of the regardlug to which slot the elliptical fr~sment speaker. "The cost?" refers. Althou~ one might suggest extensions far handling this fra~ent, a semantic Real understanding consists sot only of strategy alone does not provide an adequate frame~- resognlzAr~ the particular surface-request or wurk for Interpreting intersentential ellipsis. surface-lnform, but also of inferring what the speaker wants to accomplish and the relationship The third potential strategy utilizes a model of each utterance to this task. Interpretation of c~ the information-seeker's inferred tank-related ellipsis based upon the speaker's inferred under~ plan and discourse ~oals. The power of this lying task-related plan and discourse Eoals facil- approach is its reliance upon pragmatic informa- itates a richer interpretation of elliptical tion, including discourse content and converse- utterances. tiona~ goals, rather than upon precise representations of the preceding utterances alone. REQUISITE KNCWLEDG E Allen(1980) was the first to relate ellipsis processlug to the domain-dependent plan underlying A speaker can felicitously employ intersen- a speaker's utterance. Allen views the speaker's tentlal ellipsis only Lf he believes his utterance utterance as part of a plan which the speaker has will be properly understood. The motivation for constructed and is executlug to accomplish his this work is the hypothesis that speaker and overall task-related goals. To interpret ellipti- hearer mutually believe that certain knowledge has cal fragments, Allen first constructs a set of been acquired during the course of the dialogue possible surface speech act representations for and that this factual knowledge along with other the elliptical fragment, limited by syntactic processing knowledge will be used to deduce the clues appearing within the fragment. The task- speaker,s intentions. We claim that the requisite related ~oals which the speaker might pursue form factual knowledge includes the speaker,s inferred a set o1" expectations, and Allen attempts to infer task-related plan, the speaker's inferred beliefs, the speaker's ~al-related plan which resulted in and the anticipated discourse Eoala of the execution of the observed utterance. A part of speaker; We claim that the requisite processing this inference process involves determining which knowledge includes plan recognltlon strategies and of the partially constructed plans connecting focuslng techniques. expectations (goals) and obeerved utterance are 'reasonable given the knovled~ and mutual beliefs 1. Task-Related Plan of the speaker and hearer. Allen selects the surface speech act which produced the most reasonable In a cooperative information-seeking dAelo- inferred plan as the correct interpretation. gue, the ln~ormation-provider is expected to infer the ir~ors~ation-seeker, s underlying task-related Allen notes that the speaker's fragment must plan an the dialogue pro~-eases. At any point An identif7 the subg~als which the spea~er is pursu- the dialo~e, ZS (the information-seeker) believes Lug, but claims that in very restricted dmaains, that soae subset of this plan has been coemunA- identifying the speaker's overall ~ from the mated to IP (the in~ormation-provider); therefore utterance ls sufficient to identify the appropri- Y~ feeAa Juatl.rled in ~ormuAating utterances under ate response in terms of the obstacles present in the assumption that IP will use this inferred task such a plan. For his restricted do~aln involving model to interpret utterances, includIDg elliptl- train arrivals and departures, Allen's Interprets- eLL frasmente.

189 An example will illustrate the importance of regarding IS at the time the elliptical fragment IS's inferred task-related plan in interpreting is uttered. In the latter case, IP believes it As ellipsis. In the following, IS is conslderi~ mutually believed that IS already knows IP' s purchase of a home mentioned earlier in the dialo- beliefs regarcling when C/~O0 meets, so a request ~ue: for that informatlon is not felicitous and a different intention or discourse goal is attributed IS: "What elementary school do children to L~. in Rolling Hills attend?" Allen and Perrault(1980) used mutual beliefs ZP: "They attend Castle Elementary." in their work on indirect speech acts and sug- ~sted their use in clarification and correction IS: "Any nearby seim clubs?" dlalogues. ~idner(1983) models user beliefs about system capabilities in her work on recognlzlng An informal poll indicates that most people inter- speaker intention in utterances. pret the last utterance as a request for swim clubs near the property under consideration in 3. Anticipated Discourse Goals Rolling Hills and that the reason for such an interpretation is their inference that IS is The speaker' s anticipated discourse goals investigating recreational facilities that might form a third compocent of factual knowledge be used if IS were to purchase the home. However, required for processing elliptical frasmenta. The if we substitute the frasment dlalogue precedlng an elliptical utterance may •An~ nearby day-care centers?" sugEest discourse goals for the speaker; these for the last utterance in the dialogue, then sugEested discourse gcals become shared knowledge interpretation depen~ upon whether one believes between speaker and hearer. As a result, the IS wants hls/her children to be bused, or perhaps listener is on the lookout for the speaker to pur- even walk, to day-care directly from school. sue these anticipated discourse goals and inter~ ~rets utterances accordingly. 2. Shared Beliefs Consider for example the following dialogue: Shared beliefs of facts, beliefs which the listener believes speaker and iistecer mutually IP: "Have you taken C3105 or C3170?" hold, are a second component of factual knowledge required for processing intersentential elliptical I~: wit the Unlversity of Delaware?" fra6ments. These shared beliefs either represent presueed a priori knowledge of the domain, such as IP: "No, anywhere." a pres~ptlon that dialogue participants in a unAvereity domain know that each course has a IS: "Yes, at Penn State." teacher, or beliefs derived from the dialogue itself. An e~ple of the latter occurs i~ IP In this example, IP's inlt~al query produces a tells IS that C3360 is a 5 credit hour course; IS strong anticipation that IS will pursue the may not himself believe that C3360 is a 5 credit discourse 8oal of provldlng the requested i~forma- hour course, but as a result of IP's utterance, he tlon. There/ore subsequent utterances are inter- does believe it is mutually believed that IP preted with the expectation that IS will eventu- believes this. ally address this 8oal. IS's first utterance is interpreted as ~u-sulng a discourse Eoal of seek- Understanding utterances requires that we ing clarification of the question posed by IP; identify the speaker's discourse goal in making IS' s last utterance ansMers the initial query the utterance. Shared beliefs, often called posed by IP. However discourse expectatlons do mutual beliefs, form a part of communicated not persist forever with intervening utterances. knowledge used to interpret utterances and identify discourse goals in a cooperative dlalogue. . Processing ~owledp The following e~a~le illustrates how IP' s beliefs about IS influence usderstan~Ing. P1 an- recognl tlon strategies and focusing techniques are necessary components of processing IS: "Who is teaching C~O0?" knowledge for interpreting intersententlal eillpsis. Plan-recognltion strategies are essen- IP: "Dr. Brown is teaching C.~O0." tial I- order to In/er a model of the speaker's underlying task-related plan and focusing tech- IS: "At ni~t?" niqces are necessary in order to identIDi that portion of the underlying plan to which a frasmen- The frasmentar~ utterance "At ni~t?" is a request tar7 utterance refers. to know whether CS~O0 is meeting at night. Hc~- ever, if one precedes the above utterances with a Focusing mechanAas have been employed by quer~ whose rms~onse informs IS that CS~O0 meets Gross(1977) in identifying the referents of defin- only at ni~t, then the last utterance, ite noun phrases, by Robinson(1981) in interpret- •At ni~t? = ing verb p~vases, by ~ner( 1981 ) in anaphora becomes an objection and request for corroboration resolution, by CarberrT(1983) in plan inference, or e~lanatlon. The reason for this difference in and by McKeown(19fl~) in natural lan&uage genera- interpretation is the difference in beliefs t~on.

190 FRAmeWORK FOR PROCESSING ELLIPSLS Recourse to such a belief model is necessary in order to allow for Yes-No questions to which If an utterance is parsed as a sentence frag- the answer is "No" and yet eliminate potential ment, ellipsis processing begins. A model of any associations which a human listener would reCOg- preceding dialogue contains a context tree (Car- nize as unlikely. Although this discarding of berry, 1983) corresponding to IS's inferred under- possible associations does not occur often in lying task-related plan, a space containing IS's interpreting elliptical fragments, actual human anticipated discourse goals, and a belief model dialogues indicate that it is a real phenomenon. representing IS's inferred beliefs. (Sidner(1981) employs a similar strategy in her work on anaphora resolution. A co-specifler pro- Our framework is a top-down strategy which posed by the focusing rules must be confirmed by uses the informatlon-seeker' s anticipated an inference machine; if any contradications are discourse goals to guide interpretation of the detected, other co-specifiers are suggested. ) fragment and relate it to the underlying task- related plan. The discourse component first A propositional fragment can be of two types. analyzes the top element of the discourse stack The first contains a proposition whose name is the and suggests potential discourse goals which IS same as the name of a proposition in the plan might be expected to pursue. The plan analysis domain. The second type is a more general propo- component uses the context tree and the belief sitional fragment which cannot be associated with model to suggest possible associations of the a specific plan-based proposition until after elliptical fragment with aspects of IS's inferred analyzing the relevant propositions appearing in task-related plan. If multiple associations are IS's plan. The semantic representations of the" suggested, the evaluation component applies utterances focusing strategies to select the interpretation believed intended by the speaker --- namely, that "Taught by Dr. Smith?" most appropriate to the current focus of attention "With Dr. Smith?" in the dialogue. The discourse component then uses the results produced by the analysis com- would produce respectively the type I and type 2 ponent to determine if the fragment accomplishes pro pc si ti ons the proposed discourse goal; if so, it interprets the fragment relevant to the identified discourse Teaches (_as :&SECTIONS, SMITH ) goal. Genpred( SMITH ) The latter indicates that the name of the specific PLAN-ANALYSIS COMPONENT plan proposition is as yet unknown but that one of its parameters must associate with the constant I. Association of Fragments Sml th.

The plan-analysls component is responsible A proposition of the first type associates for associating an elliptical fragment with a term with a proposition of the same name if the parame- or conjunction of propositions in Is's underlying ters of the propositions associate. A proposition task-related plan. The analysis component deter- of the second type associates with any proposition mines, based upon the .current focus of attention, whose ~arameters include terms associating with the particular aspect of the plan highlighted by the known parameters of the propositional frag- IS's fragment and the discourse goal rules infer ment. hcw IS intends the fra@Rent to be interpreted. This paper will discuss three classes of ellipti- The semantic representation of a term such as "The meeting time?" cal fragments; a description of how other frag- is a variable term ments are associated with plan elements is pro- _~me :&MTG- TMES vided in (Carberry, 1985). Such a term associates with terms of the same semantic type in IS's plan. Note that the exlst- A constant fragment can only associate with ing plan may contain constant instantiatlons in terms whose semantic type is the same or a super- set of the semantic type of the constant. Further- place of former variaOles. A term fragment still more, each term has a limited set of valid instan- associates with such constant terms. tlations within the existing plan. A constant associates with a term only if IP's beliefs indi- 2. Results of Plan-Analysis Component cate that IS might believe that the uttered constant is one of the te.,-m's valid instantiations. The plan-analysis component constructs a con- term For example, if a plan contains the proposition junction of propositions PLPREDS and/or a Starting-Date( AI-CONF, JAN/5) PLTERM representing that aspect of the the elliptical fragment informatlon-seeker' s plan highlighted by the •February 2?" elliptical fragment; STERM and SPREDS are produced wall associate w~th this proposition only if IP by substituting into PLTERM and PLPREDS the terms believes I3 might believe that the starting date in IS's fragment for the terms with which they are for the AS conference is in February. associated in IS's plan.

191 (1)mEarn-Credit(IS,CS360,FALL85) such that Course-Offered(CS360,FALL85) ] i (1)~Earn-Cre~t-Sectlon(IS,_ss:&SECTIONS) such that Is- ~ection-Of(_ss: &3ECTION S, ~360 ) Is- Of fere,~(_ss: &SECTION S, FALL85 )

(1)~iearn-Materlal(IS,_ss:&SEcTIONS,_s~l:&SYLBI) such that Is-Syllabus-Of(_ss:&SECTIONS,_s~l:&SYLBI)

i J (1)ILearn-Frem(I~,_fac:aSECTIONS,_ss:&SECTIONS) (1)'iearn-Text(IS,_txt:&TEXTS) such that such that Teaches(_fae:&FACULTY,_ss:&SECTIONS) Uses(_ss:&SECTIONS,_txt:&TEXTS) [ i (1)IAttend-CIass(IS,_day:&MTG-DAYS,_tme:&MTG-T~S,~Ic:&MTG-PLC3) such that Is- Mt g-Day (_ss: &SECTION S, day: &MTG- T~S ) Is-Mtg-Time (_ss: &SECT ION S,_tme: &~- T~S ) Is-Mtg-PIc(_ss:&SECTIONS,_plc:&MTG-~C~)

Figure I: A Portion of the Expanded Context Tree for EXAM~E- I

It appears that h,-,ans retain as much of the If the elliptical fragment is a constant, term, or established context as possible in interpreting term with attached propositions, the analysis com- intersententlal ellipsis. Carbonell(1983) demon- ponent produces a term STERM associated with the strated this phemonenon in an informal poll in constant or term in the fraRment as well as a con- which users were found to interpret the fraRment Junction of propositions SPREDS. SPREDS consists in the followlng dialogue as retaining the fixed of all propositions along the paths from the root media specification: of the context tree to the nodes at which an element of the frasment is associated with a plan "What is the size of the 3 largest element, as well as all propositions appearing single port fixed media disks?" along the previous ACTIVE path. The former represent the new context derived from IS's frs4- "disks with two ports?" mentary utterance whereas the latter retain the previously established global context. We have noted the same phenomenon in a student advisement domain. 3. E~mple

Thus when an elliptical fragment associates This example illustrates how the plan- with a portion of the task-related plan or an analysis component determines that aspect of IS's expansion of one of its actions, the context esta- plan hi~llg~ted by an elliptical fragment. It bllshed by the preceding dlalogue must be used to also shows how the established context is main- replace information deleted from this streamlined, rained in interpreting ellipsis. frae~mentary utterance. The set of ACTIVE nodes in the context model form a stack of plans, the toP- IS: "Is C3360 offered in Fall 1985?" most of whlca is the current focused plan; each of these plans is the expanslon of an action IP: "Yes." appearing in the plan Immediately beneath it in this stack. These ACTIVE nodes represent the IS: sod any sections meet on Monday?" established Elobal context within w~ich the frag- mentary utterance occurs, and the propositions IP: "One section of CS360 meets on Monday at ~PM appeaclng along this path contain information and another section meets on Monday at 7PM. " missing frca the sentence fragment but ;~'esumed understood by the speaker. IS: "The text?"

If the elliptical fragment ls a proposition, A portlon 0£ I~'s inferred task-related plan prior the analysis component produces a conjunction of to the elliptical fragment is shown in glgure I. propositions 3PREI~ representing that aspect ot Nodes along the ACTIVE path are marked by aster- the plan hi~hii~ted bY IS's el!iptlcal fra~ent. lsk~.

192 The semantic representation of the fragment when a plan is explicitly focused. For example, "The text?" the actions of learnlr~ from a particular teacher, will be the variable term learning the material in a given text, and attend- _book: &TEXTS Ing class will all reside at the same focus level This term associates with the term within the expanded plan for earning credit in a _txt :&TEXTS course. The action of going to the cashler's appearing at the node for the action office to pay one's tuition also appears within this expanded plan; however it will reside at a Learn- Text (IS, txt: &TEXTS ) different focus level since it does not come to such that mind nearly so readily when one thinks about tak- Use s(_ss: &SECTIONS,_txt :&TEXTS ) ing a course.

The propositions along the active path are The following are two of seven focusing rules used to select the association deemed most Course-Offered( CS360, FALL85 ) relevant to the existing plan context. Is- Sectl on- Of (_ss: &SECTIONS, CS360) Is- Offered (_as :&SECT I0N S, FALLS 5 ) [F1] Within the current focus space, prefer asso- Is-Syllabus-Of(_ss: &SECTIONS,_syl: &S~LBI) clatlons which occur within the current Teaches (_fac: &FACULTY,_ss: &SECTIONS) focused plan. I s- Mt g-Day (_ss: &SECT ION S, MDN DAY ) Is- Mt g-Time (_ss: &SECT IONS,_tme: & M%T,-T~S ) IF2] Within the current focus space and current Is- Mt g- P1 c (_ss: &SECT IONS,_pl c: &MTG- PLCS ) focused plan, prefer associations within the actions to achieve the most recently con- These propositions maintain the established considered action. text that we are talking about the sections of C3360 that meet on Monday in the Fall of 1985. The path from the root of the context model to the DISCOURSE GOALS node at which the elliptical fragment associates with a term in the plan produces the additional We have analyzed dialogues from several dif- pro pc sl tl on ferent domains and have identified eleven Uses (_ss :&SECT IONS,_book: &TEXTS ) discourse goals which occur during information- The analysis component returns the con~unctlon of seeking dialogues and which may be accomplished these propositions along with STERM, in this case via elliptical fragments. Three exemplary _book: &TEXTS discourse goals are The semantics of this interpretation is that IS is [;] drawing attention to the term STERM such that the Obtaln-In/ormatlon: IS requests Ir.formatlon relevant to constructing the underlying con~unctlon of propositions SPREDS is satisfied task-related plan or relevant to formulating --- namely, the textbook used in sections of C3360 an answer to a question posed by IP. that meet on Monday in the Fall of 1985. [2] Obtaln-Corroboration: IS expresses surprise EVALUATION COMPONENT regarding some proposition P and requests elaboration upon and justification of it. The analysis component proposes a set of Seek-Clarify-questlon: IS requests informa- potential associations of the elliptical fragment [33 tion relevant to clarifying a question posed with elements of IS' s underlying task-related by ZP. plan. The evaluation component employs focusing strategies to select what it believes to be the interpretation intended by 13 --- namely, that ANTICIPATED DISCOURSE GOALS interpretation most relevant to the current focus of attention in the dialogue. When IS m~es an utterance, he is attempting We employ the notion of focus domains in to accomplls~ a discourse goal ; this discourse goal may in turn predict other suDsequent order to group finely grained actions and associ- discourse goals for IS. For e~ple, if I~ asks a ated plans into more general related structures. question, one anticipates that IS may want to A focus domain consists of a set of actions, one expand upon his question. Similarly, utterances of which is an ancestor of all other actions in made by IP suggest dlsoourse goals for LS. These the focus domain and is called the root of the Aatlcipated Discourse Goals provide very strong focus domain. If as action is a member of a focus domain and that action is not the root action of expectations for IS and may often be accomplished another focus domain, then all the actions con- implicitly as well as explicitly. talnad in the plan associated with the first The discourse ~als of the previous section action are also members of the focus domain. also serve as anticipated discourse goals. Three (This is similar to Grosz's focus spaces and the appear notion of an object being in implicit focus.) additional anticipated discourse goals tO play a major role in determining how elliptical The use of focus domains allows the groupin8 fragments are interpreted. One such anticipated together of those actions that appear to be at discourse ~al is: approximately the sa~me level of Impllcit focus

193 Accept-Questlon: IP has posed a question to sug~sted discourse goal. If a discourse goal IS; IS must now accept the question either rule succeeds in producing an interpretation, then explicitly, implicitly, or indicate that he the discourse component identifies that discourse does not as yet accept it. goal and its associated interpretation as its understanding of the utterance. Normally dialogue participants accept such questions implicitly by proceding to answer the ques- I. Discourse Expectation Rules tion or to seek information relevant to formulating an answer. However IS may refuse to accept The top element of the discourse stack the question posed by IP because he does not activates the discourse expectation rule with understand It (perhaps he is unable to identify which it is associated; this rule in turn suggests some of the entities mentioned in the question) or discourse goals which the information-seeker' s because he is surprised by it. This leads to utterance may pursue and activates these discourse discourse goals such as seeking confirmation, goal rules. The following is an example of a seeking the identity of an entity, seeking clarif- discourse expectation rule: ication of the posed question, or expressing surprise at the question. [DE1]If the top element of the discourse stack is Answer-Question, then I. Apply discourse goal rule DG-Answer-Quest THE DISCOURSE STACK to determine if the elliptical fragment is being used to accomplish the discourse goal The discourse stack contains anticipated of answering the question. discourse goals which IS is expected to pursue. 2. If no interpretation is produced, apply Anticipated discourse goals are pushed onto or rule S-Suggest-Answer-Questlon to determine popped from the stack as a result of utterances if the elliptical fragment is being used to made by IS and IP. We have identified a set of accomplish the discourse goal of suggesting stack processing rules which hold for simple an answer to the question. utterances. Three examples of such stack process- 3. If no interpretation is produced, apply Ing rules are: discourse goal rule DG-Obtaln-Info to determine if the elliptical fragment is being used [SP1]When IP asks a question of IS, Answer- to accomplish the discourse goal of seeking Question and Accept-Questlon are pushed onto information in order to construct an answer the discourse stack. to the posed question.

[SP2]When IS poses a question to IP, Expand- Once IS understands the question posed to him, Question is pushed onto the discourse stack. IP's strongest expectation is that IS will answer Once IP begins answering the question, the the question; therefore first preference is given stack is popped up to and including the to interpretations which accomplis~ this goal. If Expand-Questlon discourse goal. IS does not immediately answer the question, then we expect a cooperative dialogue participant to [SP3]When IS's utterance does not pursue a goal work towards answering the question. This entails sugEested by the top entry on the discourse gathering information about the underlying task- stack, this entry is popped from the stack. related plan in order to construct a response.

The motivation for these rules is the following. 2. Discourse Goal Rules When IP asks a question of IS, IS is first expected to accept the question, either implicitly Discourse goal rules determine if an elllptl- or expllcltly, and then answer the question. Upon cal fragment accomplishes the associated discourse posing a question to ~P, IS is expected to expand goal and, if so, produce the appropriate upon this question with subsequent utterances or interpretation of the fragment. These discourse wait u~tll IP produces an answer to the question. goal rules use the plan-analysls component to help Alt~oug~ the strongest expectations are that IS determine the best interpretation of the frasmen- will pursue a goal suggested by the top element of tary utterance relevant to the suggested discourse the discourse stack, this anticipated discourse goal. However these interpretations are not goal can be passed over, at which point it no actual representations of surface speech acts; longer sug~sts expectations for utterances. instead they generally indicate elements of the plan whose values the speaker is querying or specifying. In many respects, this provides a DISCOURSE INTERPRETATIOM COMPOM~T better "understanding" of the utterance since it describes what the speaker is trying to accom- The discourse component employs discourse pli~. expectation rules and discourse goal rules. The discourse expectation rules use the discourse The following is an example of a rule associ- stack to suggest possible discourse goal s for L~ ated with a discourse goal suggested by the stack and activate the associated discourse goal rules. entry Accept-Response; the latter is pushed onto These disnourse goal rules ttse the plan-analysis the discourse stack when IP responds to a question component to help determine the best interpreta- posed by IS. tion of the fra~entar7 utterance relevant to the

194 Obtain-Corrob Acre pt- Response The discourse component calls the plan- Obtaln-Informatlon analysis component to associate the elliptical fragment with a term STERM or a conjunc- The discourse goal rules sugEested by Accept- tion of propositions SPREDS in IS's underly- Response do not identify the fragment as accom- ing task-related plan. If IP believes it is plishing their associated discourse Eoals, so the mutually believed that IS already knows IP's top entry of the discourse stack is popped; this beliefs about the value of the term STERM or indicates that IS has implicitly accepted IP' s the truth of the propositions $PREDS, then response. The entry Obtaln-Informatlon on the identify the elliptical fragment as accom- discourse stack activates the rule DG-Obtaln-In/'o. plishing the discourse ~al of expressing Pl an- analy sl s is activated to associate the surprise at the preceding response; in par- elliptical fragment with an aspect of I$'s task- tlcular, IS is surprised at the known values related plan. The construction of 5TERM and of STEP=M or SPREDS in li@~t of the new infor- SPREDS for this ezample was described in detail in met.lon provided by IP' s preceding response the plan analysis section and will not be repeated and the known aspect queried by IS's frag- here. Since our belief model indicates that IS ment. does not currently know the value of STERM such that SPREDS is satisfied, this rule identifies the The followin8 is one of several rules associ- elliptical fragment as seeking information in ated with the discourse ~al Answer-Question. order to formulate a task-related plan; in particular, I -~ is requestlng the value of STERM such J~Ct"Answer- Oues t--~. that SPREDS is satisfied --- namely, the textbook If the elliptical fragment terminates with a used in sections of C3360 that meet on Monday in period, then the discourse component calls the Fall of 1985. the plan-analysls component to associate the elliptical frasment with a conjunction of propositions SPEEDS in IS's underlying task- This example illustrates an utterance in which IS related plan. If successful, interpret the is surprised by IP's response and see~s elabora- elliptical fragment as answerlr~ "Yes", with tion and corroboration of it. (The construction the restriction that the propositions SPREDS of $PREDS by the plan analysis component will not be satlsfi~d in the underlyin~ .i ~n. be described since it is similar to EXAMPLE-I.)

IS: "I want to take CS620 in Fall 1985. Who is teaching it?" IMPLE}~NTATION AND EXAMPLES IF: "Dr. Smith is teaching CS620 in Fall 1985." This pragmatics-based framework for processing intersententlal ellipsis has been implemented IS: "What time does CS620 meet?" for a subset of discourse goals in a domain con- slstln8 of the courses, policies, and requlrements IP: "C°~20 meets at SAM. " for students at a unlverslty. The following are worklng examples from this implementation. IS: "With Dr. Smlth?"

The ellipsis processor is presented with a I~'s elliptical fragment will associate with the semantic representation of Is's elliptical frag- term ment; it "understands" intersententlal elliptical Teaches (_fat -&FACULTY,_ss :&SECTIONS ) utterances by Identlfyin8 the discourse goal which in IS's task-related plan. SPREDS will contain I~ is pursuing in employing the frasment and by the propositions producing a plar,-Oased interpretation relevant to Course- Offered( CS6 20, FALL85 ) this discourse goal. Is- Section- Of(_ss :&SECTIONS, CS620 ) Is- Offered (_ss: &SECT I0N S, gALL85 ) Is-Syllabus-Of( _ss :&ZECTIONS,_sy i :&SYLB I ) This e,-=mple illustrates a simple request for Teaches( SMITH ,_ss :&SECTIONS) information. Is-Mt~-Day (_ss: &SECTIONS,_day :&MTG-DA YS ) Is-Htg-Time(_ss: &SECTIONS,_tme: &MT~- TM~S) IS: "Is CS360 offered in Fall 19857" Is- Mtg-Plc(_ss: &SECTIONS,_gl c :&MTG- PL CS)

IP: "Yes." Immediately prior to the occurrence of the elllpt- ical fragment, the discourse stack contains the IS: "Do any sections meet on Monday?" entries

IP: "One section of C3360 meets on Monday at qPM Acre pt- Respo n~e and another section meets on Monday at 7PM. " Obtain- Information

IS: "The text?" Accept-Response, the top entry of the discourse stack, su6Eests the discourse goals of I )seeking .~onflrmatlon or 2,~seeklng corroboration of a com- Immediately prior to IS's elliptical utter- ponent of the preceding response or 3)seeking ela- . ante, the discourse stack contair~ the entries boration and corroboration of some aspect of this

195 ( I ) eEarn-Credit (IS ,_crse :&COU RsE,_sem: &SEmeSTERS) such that Course-Of f ered(_cr se: &COU RSE,_sem: &S~STERS) l I ( I ) eEarn-Cr edit-Sectlon(IS ,_ss: &SECTIONS) such that Is- Secti on- Of (_as: a3ECT ION S, _or se :&COURSE) Is-Offered(_ss: &SECTIONS,_sea: &SE)~STERS) I i ( I )iRegl ster- Late (IS ,_ss: &SECTION S, _sea: &S E)~STERS) i

I [ (2 ) eMiss- Pro- Reg( IS ,_sea: &SEM~TEBS) (2) Pay-Fee (IS, LATE- REG ,_sere: &SEI~STF~S) t [ (2) Pay( IS ,_lreg: &MONEY) such that Costs( LATE- RE3 ,_lreg: &MON~-Y)

Figure 2. A Portion of the Expanded Context Tree for EXAMPLE-3

response. The discourse goal rules Seek-Conflrm preceding the elliptical frasment. The and Seek-Identlfy fail to identify their associ- parenthesized numbers preceding actions indicate ated discourse goals as accomplished by the user's the action's focus domain. I~'s fragment associ- fragment. ates with the term _ireg: &MONEY Ou~ belief model indicates that IS already in IS' s inferred plan, as well as with terms else- knows that SPREDS is satisfied; therefore the where in the plan. However none of the other discourse goal rule DG-Obtain-Corrob identifies terms appear in the same focus space as the most the elliptical fragment as expressing surprise at recently considered action, and therefore the and requesting corroboration of IP's response. In association of the fragment with particular, IS is surprised that SPRED~ is satis- _lreg: &MONEY fied and this surprise is a result of is selected as most relevant to the current dlalogue context. The discourse stack immediately [I] the new information presented in IP's preced- prior to the elliptical fra6ment contains the sin- ing response, namely that 8AM is the value of gle entry the term _tae: &MTG- T~S Prov ide- For- Assimil atl on in the SPREDS proposition This anticipated discourse goal suggests the Is- Mt g-Tiae(_ss: &SECTION S,_tme: ~ T~S ) discourse goals of 1 )providing further inforaatlon for assimilation and 2)see~Ing information in C2] the aspect of the plan queried by IS's order to formulate the task-related plan. The elliptical fra~ent, namely the SPREDS propo- utterance terminates in a "?", ruling out provide sition for assimilation. Therefore rule DG-Obtaln-Info Teaches ( SMITH ,_ss: &SECTIONS) identifies the elliptical fragment as seeking information. In particular, the user is request- EXA~ELFcl ing the fee for late registration, namely, the The following is an example which our framework value of the term handles but which poses problems for other stra- _cstl :&MONEY te61es. such that SPREDS is satisfied, where SPREDS Is the conjunction of the propositions IS: "I want to register for a course. But I massed pre-reglstration. Course-Offered(_crs: &COU RSE,_sea: &SEMESTERS ) The cost?" Is-Sectlon-Of( _ss: &SECTION S,_sem: &SE)~STERS) Is- Offer ed(_ss: &SECTIONS,_sem: &SEmeSTERS) The first two utterances establish a plan context Costs( LATE- Rwn. ,_cstl :&MONEY) of late-reglstering, within which the elliptical fra~ent requests the fees involved in doing so. ( Late registration generally involves extra chargos. )

Figure 2 presents a portion of 13' s underlying task-related plan Inferred frca the utterances

196 EXTENSIONS AND FUTURE WORK REFERENCES I. Allen, J.F. and Perraul t, C.R. , "Analyzing The main limitation of this pragmatics-based Intention in Utterance, s", Artificial Intelli- framework appears to be in handling Intersenten- gence, 15(3), 1980 tlal elliptical utterances such as the following: 2. Carberry, S., "Tracking User Goals in an Informatlon-Seeking Environment", AAAI, 1983 IS: "Who is the teacher of C3200?" 3. Carberry, S., "Pragmatic Modeling in Informa- tion System Interfaces", forthcoming Ph.D. IF: "Dr. Herd is the teacher of C3200." Dissertation, Dept. of Computer Science, University of Delaware, Newark, Delaware IS: "C32637" 4. Carbonell, J.G., and Philip Hayes, "Recovery Strategies for Parsing Extragrammatl cal Obviously IS' s elliptical fragment requests the Language", Amer. Journal of Comp. Ling. , teacher of C3263. Our model cannot currently han- Vcl.9, No.3-4, 1983 dle such fragments. This limitation is partially 5. Carbonell,J.G., " "Discourse Pragmatlcs and due to the fact that our mechanlems for retaining Ellipsis Resolution in Task-Orlented Natural dialogue context are based upon the view that IS Language Interfaces" , Proc. 21 rst Annual constructs a plan for a task in a deptb-flrst Meeting of ACL, 1983 fashlon, completing Investlgation of a plan for 6. Flowers, M. and M.E. Dyer, "Really Arguing C3200 before moving on to investigate a plan for With Your Computer", Proc. of Nat. Comp. CS263. Since the teacher of C3200 has nothing to Conf. , 1984 do with the plan for taking C3263, the mechanisms 7. Grice, H.P., "Meaning", Phil. Rev. 66, 1957 for retaining dialogue context will fail to iden- 8. Orice, H.P., "Utterer's Meaning and Inten- tify • teacher of CS263" as the information tions", Phil. Rev., 68, 1969 requested by IS. 9. Grosz,B.J., "The Representation and Use of Focus in a System for Understanding Dialogs", One might argue that the elliptical fragment IjCAI, 1977 in the above dialogue relies heavily upon the syn- 10. Orosz,H.J., Joshi, A.K., and Weinstein, S., tactic representation of the preceding utterance •Providing a Unified Account of Definite Noun and thus a syntactic strategy is required for Phrases in Discourse ", Proceedings 2 Irst interpretation. This may be true. However if we Annual Meeting of ACL, 1983 view dialogues such as the above as investigating 11. Hendrlx, G.G., E.D. Sacerdoti, and J.Slocum, task-related plans in a kind of "breadth-flrst" "Developing a Natural Language Interface to fa~hlon, then IS is analyzing the teachers of each Complex Data", SRI International, 1976 course under consideration first, and will then 12. hltman,D.J., and Allen, J.F., "A Plan Recog- move to considering other attributes of the nation Model for Clarification Subdlalogues", courses. It appears that the plan-based framework Proceedings of the International Conference can be extended to handle many such dialogues, on Computational Linguistics, 198~ perhaps by using meta-plans to represent how IS is 13. Mann,W., J.Moore, and J.Levin, "A Comprehen- constructing his task-related plan. sion Model :'or Human Dialogue", IJCAI, 1977 14. McKeown,K. R., "The Text System /or Natural Language Generation: An Overview", Proc. of CON CL USION S the 20th Annual Meeting of ACL, 1982 15. Perrault, C.R., and Allen, J.F., "A Plan- This paper ha~ described a pragmatlcs-based Based Analysis of ~ndlrect Speech Acts", approach to interpreting intersententlal ellipti- American Journal of Computational Llnguls- cal utterances during an information-seeking tiCS, July 1980 dialogue in a task domsin. Our framework coordl- 16. Robinson, A. E., "Determining Verb Phrase nares many knowledge sources, including the Referents in Dialog~", American Journal of informatlon-seeker' s inferred task-related plan, Computatlor~l Linguistics", Jan. 1981 his inferred beliefs, his anticipated discourse 17. Sidner, C.L., "What the Speaker Means: The goals, and focusing strategies to produce a rich Recog~%itlon Of Speakers' Plans in Discourse", interpretation of ellipsis, including identifica- Comp. and Maths. with Appls., Vol.9,No. 1, tion of the Ir~ormatlon-seeker's d/scourse goal. 1983 This framework can handle many e-~mples wblch pose 18. Si~ner,C.L., "Focussing for Interpretation of problems for other strate~Les. We claim that the Pronouns", American Journal of Computational advantage of tbls approach is its reliance upon Linguistics, Oct. 1981 pragmatic information, including discourse content 19. Waltz, D.L., "An Engllsh Language Question and conversational goals, rather than upon precise An~werlng System for a Large Relational Data representations of the preceding utterance alone. Base", Comm. of ACM, vo121, No.7, 1978 20. Webber, B.L., M.E. Pollack, and J. Hirsch- berg, "User Participation in the Reasoning ACKN OWLEDG E~ TS Processes of Expert Systems", Proc. of Nat. Con/. on Art. Int., 1982 T would llke to thank Ralph Welschedel for 21. Welsehedel, R.M. and N. Sondhelmer, "An his encouragement and direction in this research Improved Heuristic for Ellipsis Processing", and Lance Remsbaw for many help/ul ~Iscusslons and Proc. 20th Annual Meeting of ACL, 1982 suggestlons.

197