Dialogue Acts, Synchronizing Units, and Anaphora Resolution

In this paper, we present the results of a corpus analysis, and a model of anaphora resolution in spontaneous spoken dialogues. The main finding of our corpus analysis is that less than half the pronouns and demonstratives have NP antecedents in the preceding text; 22% have sentential antecedents and the remainder have no identifiable linguistic antecedents. As part of the corpus analysis we present the results of inter-annotator agreement tests. These were carried out for the annotation of anaphor types and their antecedents, and for the segmentation of the dialogues into dialogue acts. The results of the inter-annotator agreement tests indicate that our classification method is reliable and that the annotated dialogues can be used as a standard against which to measure the performance of the anaphor resolution algorithm. The algorithm, based on Strube (1998), is capable of classifying pronouns and demonstratives, and co-indexing anaphors with NP and sentential antecedents. The domain from which potential antecedents for both individual and discourse-deictic anaphors can be elicited is defined in terms of dialogue acts. The dialogue segmentation method uses dialogue acts to form Synchronizing Units, which reflect the achievement of common ground (Stalnaker 1974, 1979). We show that predicate information, NP form, and dialogue structure can be successfully used in the resolution process. 1 I N T R O D U C T I O N In this paper, we present a model for the resolution of pronouns and demonstratives in spontaneous spoken dialogue. In the semantic, syntactic, and psycholinguistic literature, work on anaphora has concentrated primarily on the analysis of pronouns and definite NPs with NPantecedents. This is considered to be the 'normal' type of anaphoric reference. Our corpus study reveals that in actual language use this type of anaphoric reference accounts for less than half of the occurrences of pronouns and demonstratives (45%). An additional 22% are anaphors with sentential and VP-antecedents. Although this type has been studied previously (Webber 1991 and, particularly, Asher 1993 provide extensive theoretical accounts), it seems that its frequency and therefore importance has been largely underestimated. Rather surprisingly also, the remaining 52 Dialogue Acts, Synchronizing Units, and Anaphora Resolution third of all pronouns do not have identifiable linguistic antecedents of any kind. These are pronouns that are used to refer to inferrable entities and those that are used to refer to a vaguely defined general discourse topic. These findings indicate that an important function of pronouns, aside from anaphoric reference, is that they allow the speaker to leave certain referents underspecified. In spontaneous spoken language it is simply not necessary for the participants to be able to unambiguously identify a specific referent at all times. If they fail to understand an utterance and consider avoidance of misunderstanding to be important, they can immediately request clarification—an option not available in the written medium. Furthermore, the optional use of vague pronouns greatly facilitates the task of the speaker in on-line language production. We present a model that shows how pronouns and demonstratives can be classified and, if appropriate, co-indexed with the correct antecedents. The model makes use of the surface form of the anaphor, its predicative context, and the structure of the discourse. It also presents a basis for further empirical evaluations of theoretical issues in anaphora resolution. Furthermore, we believe that it provides an important starting point for spokenlanguage resolution algorithms in the field of computational linguistics, which have so far almost exclusively dealt with anaphora in written texts. In computational linguistics, most anaphora resolution algorithms are designed to deal with the predominant type of anaphoric reference found in written texts, which involves the co-indexing relations between anaphors and NP-antecedents. Aside from the different types of anaphors found in spoken language, the structure of dialogues is less clear than the structure of written texts, with lack of punctuation and paragraphs, and many syntactically incomplete clauses making it difficult to formally define the domain for potential antecedents. For these reasons, applying existing anaphora resolution algorithms to dialogues would result in a poor performance. Our model is presented in the form of a major extension of the anaphora resolution algorithm described in Strube (1998). The Strube (1998) algorithm consists of an ordered list of salient discourse entities (S-List), which provides preferences for the antecedents of pronouns. The main characteristic of the algorithm is that preferences for intraand intersentential pronouns are dealt with in a unified manner as the update of the S-List and the anaphora resolution are performed incrementally. Essential to the success of the algorithm presented in this paper is the interaction between the identification and resolution of different types of anaphors and the determination of the domain of possible antecedents. We use dialogue act units (derived from speech acts) to provide the structure necessary for the determination of the antecedent domain and also to function as antecedents for anaphors with sentential antecedents. Miriam Eckert and Michael Strube 5 3 The paper is structured as follows: section 2 describes the theoretical observations which are important for our analysis and which have partly been incorporated into the algorithm. Section 3 describes the spokenlanguage corpus used for our empirical analysis of anaphor types and for testing the algorithm. Section 4 gives an overview of our classification system for the different types of pronouns and demonstratives we identified in the spoken dialogues. Section 5 describes how we use dialogue acts to model the establishment of common ground and to define the domain of possible antecedents for the anaphors. Our resolution algorithm is presented in section 6. Section 7 gives the results of the empirical analysis. This consists of two parts: first, we evaluated the classification system in terms of inter-annotator agreement. We deemed this step necessary in order to verify the consistency of our classification. Second, we evaluated the algorithm by applying it to the hand-annotated dialogues. Finally, sections 8 and 9 provide comparisons to related work, suggest future additions and applications of our model, and present the conclusions. 2 THEORETICAL ISSUES In this section, we present some of the issues in theoretical linguistics which we consider to be important for the process of anaphora resolution in spoken dialogue. The value of these issues has so far been expressed in theoretical terms. We consider one of the contributions of our resolution algorithm to be that it opens the possibility of testing their value empirically. 2.1 Reference and the discourse model We assume that a conversation has a model of the discourse associated with it, which is distinct from both the real world and from the syntactic representation of the discourse. Such models have frequently been described in the literature, e.g. common (background (Stalnaker 1974, 1979), discourse model (Webber 1979), files (Heim 1982), attentional state (Grosz & Sidner 1986), DRSs (Kamp & Reyle 1993). These proposed models differ in a number of important ways, such as whether they are said to exist at the semantic level (files, DRSs), the pragmatic level (Stalnaker's common ground), or the discourse level (Webber's discourse model, Grosz & Sidner's attentional state). Also, some models are proposed to represent properties of the conversational participants (Stalnaker's pragmatic presuppositions constituting 54 Dialogue Acts, Synchronizing Units, and Anaphora Resolution the common ground), whilst others represent properties of the discourse itself (DRSs, attentional state). These versions of the discourse model have in common that they contain representations of the objects that have been referred to in the discourse, known as the discourse referents (Karttunen igj6),file cards (Heim 1982) or discourse entities (Webber 1979; Kamp & Reyle 1993). The discourse model also contains the attributes of the discourse entities and the relations holding between them but for the moment we will focus only on the entities introduced by NPs in the discourse. The discourse model contains representations of the entities that are salient to both participants at a given point in the discourse because they have been referred to in the previous discourse. Using terminology from Stalnaker (1979) and Clark & Schaefer (1989), we will call the part of the model containing representations of these entities the common ground. The update of the discourse model has been the subject of considerable debate. One issue is the question of when and how entities enter into the common ground. Because conversations involve more than one participant, merely uttering a sentence does not mean that the entities referred to have entered into the common ground. It is possible, for example, for one speaker to ignore the utterance of another. Conversational participants have a number of ways in which to signal understanding of an utterance, including nods of the head, relevant further contributions to the discourse, and simple backchannels (e.g. u-huh, yeah, mmhm). In our model, if an utterance is not acknowledged by the other participant, its discourse entities are not retained in the common ground. This issue is explained in more detail in section 5. There has also been disagreement concerning the influence of NP form on update, that is, whether indefinite NPs, definite NPs and pronouns serve to update the discourse model in the same way or whether different mechanisms need to be postulated. In Russell's view (Russell 1905), indefinite NPs are not referring expressions, but rather function much like existential operators, by declaring that the set of entities described by the NP is not null. This view was subsequently challenged because it does not explain the capacity of indefinite NPs to function as antecedents of anaphoric pronouns (Grice 1975; Kripke 1979; Lewis 1979). In Heim's file change semantics (Heim 1982), the approach is taken that indefinite NPs are used to introduce new entities (file cards) to the discourse model, whereas definite NPs make use of familiar ones. A concern with making a categorical distinction between definites as NPs specifying given entities, and indefinites as NPs specifying new entities, is that there are many counterexamples in which definites are used to refer to discourse-new entities (Prince 1981). In fact, recent empirical research has Miriam Eckert and Michael Strube 5 5 indicated that the numbers are by no means negligible. Poesio & Vieira (1998) show that in their corpus 50% of definites are discourse-new. The reason is that, as noted in Prince (1981, 1982), the status of entities is far more complex than can be determined by the distinction given-new. The following are examples of the categories of discourse entities defined in Prince (1981: 233, ex. 22; 237, exx. 25-27): (1) Brand-new: I bought a beautiful dress. (2) Brand-new anchored: A guy I work with says he knows your sister. (3) Unused: Noam Chomsky went to Penn. (4) Inferrable: I got on a bus yesterday and the driver was drunk. (5) Containing inferrable: Hey, one of these eggs is broken! (6) Evoked: Susie went to visit her grandmother and the sweet lady was making Peking Duck. The categories are described by adding the distinction hearer-old-hearernew to the discourse-old-discourse-new factor. Discourse-old/new describes the information status of an entity with respect to the discourse. Hearer-old/new describes the status with respect to the hearer. A definite NP such as Noam Chomsky in (3), for example, can be discourse-new if its referent has not been mentioned before, but hearer-old because it is familiar to the addressee. Prince describes this category as unused. A discourse-new entity can be anchored by a hearer-old or discourse-old entity, as in (2), where the indefinite NP is anchored by the first person pronoun I. Inferrable entities are hearer-new, discourse-new, but 'depend upon beliefs assumed to be hearer-old, where these beliefs crucially involve some trigger entity' (Prince 1992: 309). A trigger entity can be the referent of a previously mentioned NP, as in (4), where the NP a bus, once established in the discourse, allows one to refer to expected or related entities such as the driver with a definite NP. This phenomenon is also described in Lewis (1979) as accommodation. With containing inferrables, as in (5), an NP is inferred from another NP inside it (e.g. one of these eggs from these eggs). Finally, textually and situationally evoked entities are entities that are already in the discourse model. An example of this is the referent of the sweet lady in (6), which is textually evoked by the NP her grandmother. The discourse model is not intended to reflect which entities are familiar to the hearer but rather which entities are salient at that point in the discourse. We therefore assume that indefinite and definite NPs can add entities to the discourse model because they can both cause a referent to become salient in the discourse. The category inferrable is only accounted for in certain restricted cases (discussed below). We are interested here in pronouns and demonstratives. With a few exceptions, inferrables cannot be referred to with pronouns or demonstratives unless they have previously 56 Dialogue Acts, Synchronizing Units, and Anaphora Resolution been referred to with a full NP. For the purposes of our model, an NP such as the driver should be used to introduce an entity into the discourse model in the same way as the NP a bus. In the algorithm presented here, we will make use of the notion of discourse model in order to simulate pronoun and demonstrative resolution. We do not intend to present a comprehensive model of the discourse. Our simplified model consists only of a list containing representations of the objects that have been referred to in the discourse with NPs. It is similar to Grosz & Sidner's attentional state as it is intended to contain representations of entities which are salient to the participants. We will use Webber's terminology and call these representations discourse entities (Webber 1979). The list is called S(alience)-list as the entities are ordered according to how salient they are in the discourse. The algorithm resolves pronouns by co-indexing them with the highest-ranked compatible entity in the S-list. The list in our model spans more than one utterance and is incrementally updated as the discourse progresses. This means that an entity is available for subsequent anaphoric reference as soon the NP is uttered. The model does therefore not require different mechanisms for interand intra-sentential anaphora. The details of the S-list and the resolution process are described in section 6. We first turn to other linguistic issues. 2.2 Predicate information If we say that the referent of an NP is introduced into the discourse model at the point when the NP is uttered, we can assume that from that point on the entity in the discourse model is available for subsequent anaphoric reference. We will call anaphoric reference involving NP antecedents individual anaphora. However, anaphoric reference also occurs with sentential and VP-antecedents (Webber 1991; Asher 1993). Following Webber (1991), we will call this type discourse-deictic reference. In these cases, the determination of the referent seems more complex. As can be seen from the following examples taken from Asher (1993) (his numbering in parentheses), anaphors can pick up different kinds of abstract objects such as events, states, concepts, propositions or facts specified by previous clausal constituents: (7) Event: John kicked, Sam on Monday, and it, hurt. (35 (55)) (8) Concept: Somebody [had to take out the garbage,], and Bill did it,. (246 (29)) Miriam Eckert and Michael Strube 57 (9) State: John didn't know, the answer to the problem. This, lasted until the teacher did the solution on the board. (53 (85-b)) (10) Fact: Mary proved [that the defendant was lying about the President's ignorance of the cover-up.], This, shows that the cover-up is much larger than previously thought. (245 (28.a)) (n) Proposition: The 'liberation' of the village had been bloody. [Some of the Marines had gone crazy and killed some innocent villagers. To cover up the 'mistake', the rest of the squad had torched the village, and the lieutenant called in an air strike.], At first the battalion commander hadn't believed it,. (49 (82)) Asher states that the type of referent is determined by the predicative context of the anaphor. For example, a discourse-deictic anaphor in the subject position of the intransitive verb hurt must specify an event (example 7 above), whereas an anaphor in the object position of the verb believe specifies a proposition (example 11 above). In our model, we make use of the predicative context of the anaphor to determine the type of its referent and to help distinguish between individual and discourse-deitic anaphors. For example, it is generally the case that the constituent in the object position of verbs such as assume or believe specifies an abstract entity and should therefore be co-indexed with a clause. Conversely, the constituent in the object position of the verb eat specifies a concrete entitity and should therefore be co-indexed with an NP. It is clear that such a distinction is very simplistic. For example, although the constituent in the object position of believe must specifies a proposition, and propositions are generally specified by whole clauses, this is not always the case. Certain NPs can specify abstract objects in the same way that clauses do (e.g. Jane told me [a story];. I didn't believe /(,.) Future work should therefore make use of semantic tagging of NPs to supply information such as whether their referents are abstract or concrete. However, this is a difficult task for numerous reasons. One issue, for example, is that an NP may in certain cases be used to indirectly refer to an abstract object even though it generally specifies a concrete entity. In the sentence / don't believe Jane, the NP Jane stands for some/all proposition^) expressed by Jane. Another issue that requires a more complex solution concerns reference to events that are inferrable but not explicitly mentioned, e.g.: (12) We just got back from France. It was great fun. The pronoun it specifies the event of being in France. However, the VP in 58 Dialogue Acts, Synchronizing Units, and Anaphora Resolution the preceding context specifies the event of getting back from France. Getting back implies having been in a place, so the appropriate referent of the pronoun it is available to the listener as a result of world knowledge. In the work presented here, we put these complex issues aside for the time being and use the predicate of the anaphor only as one of the features guiding the simplified anaphor classification and resolution. 2.3 Referent coercion The predicative context of the anaphor is important even when the antecedent constituent has been determined because the precise referent must still be identified. Webber (1983: 332) points out that the same text string can give rise to a variety of entities available for subsequent anaphora: (13) The Rhodesian ridgeback down the block bit me yesterday. (a) It's really a vicious beast. (b) They're really vicious beasts. In continuation (a) the singular pronoun is used to refer to the individual dog, whereas in (b) the plural pronoun references the set of dogs of that particular breed. In both instances, the textstring the Rhodesian ridgeback (modified by the PP down the block in version (a)) is used to provide the referent of the pronoun. The same variety of potential referents can be found with clausal antecedents. For example, the clause in (14) can make available an event, concept, fact, or proposition as a referent for subsequent anaphors: (14) [John [crashed the car]y],. (a) This, annoyed his parents, (event) (b) Jane did that,, too. (concept) (c) This, shows how careless he is. (fact) (d) His girlfriend couldn't believe it,, (proposition) Furthermore, Moens & Steedman (1988) provide an analysis of events that divides the event-complex into a. preparatory process, culmination and consequent state. Their analysis of adverbials shows that reference can be made to any one of these subparts of the event, as can be seen in the following example taken from Ritchie (1979), cited in Moens & Steedman (1988, ex. 1): (15) When they built the 39th Street bridge . . . (a) . . . a local architect drew up the plans. (b) . . . they used the best materials. (c) . . . they solved most of their traffic problems. Miriam Eckert and Michael Strube 59 The event building the bridge consists of a preparatory process of building the bridge, which includes the architect drawing up the plans, a culmination, which involves using the best materials, and a consequent state, which involves the solution of the traffic problems. The adverbial clause supplies the necessary subparts of the event for the alternative continuations. Instead of assuming that all levels of abstract objects and all their subparts are introduced to the discourse model by the clause that makes them available, it has been suggested that discourse-deictic reference involves referent coercion (Dahl & Hellman 1995) or ostension (Webber 1991). That is, in a process similar to accommodation (Lewis 1979), the anaphor itself is used to create a new referent in the discourse model. This means that the referents of discourse-deictic anaphors do not exist in the discourse model unless anaphorically referred to. Webber suggests that for each context there are discourse entities that stand proxy for its propositional content. Discourse-deictic anaphora involves a referring function that yields a discourse entity proposition, event, event type or state from the proxy entity. Passonneau (1991: 69) uses the following example to show that referents of discourse-deictic anaphors are lost from the discourse model immediately unless referred to again: (16) (a) [I noticed that [Carol insisted on sewing her dresses*, from nonsynthetic fabric^.], (b) That,'s an example of how observant I am. (c) And they*, always turn out beautifully. (d) # Thaty's because she's allergic to synthetics. The discourse-deictic demonstrative in utterance (b) picks out a referent described in the main clause of the first utterance (/ noticed . . .). The discourse-deictic demonstrative in the final utterance (d), however, is not capable of doing the same thing. It cannot be used to refer to the intended referent in utterance (a) (Carol insisted • • •) because of the intervening utterance (c). At the time of the final utterance the referent of the first utterance is no longer available. Intervening utterances pose no such problem for individual anaphoric reference. The pronoun they in utterance (c) is used felicitously to refer to the referent of the NP her dresses in the first utterance, in spite of intervening utterances and anaphoric references. Note, however, that in spite of the transitory qualities of discourse-deictic entities, chains of discourse-deictic references are possible, as seen in this altered version of Passoneau's example: (17) (a) [Carol insisted on sewing her dresses from non-synthetic fabric], (b) That,'s because she's allergic to synthetics. (c) It,'s also because she hates cheap materials. 60 Dialogue Acts, Synchronizing Units, and Anaphora Resolution In (17), the referent of the first clause is available for anaphoric reference both in clauses (b) and (c). The continued reference ensures that it is not lost. 24 Choice of NP-form We now turn to the differences between pronouns and demonstratives as we are interested in building a resolution algorithm for both of these NP forms. Gundel et al. (1993), amongst others, note that there is a correlation between different NP forms and the accessibility of their referents. Pronouns and demonstratives provide only little information concerning the identity of their referents (in English, number and gender only) and are therefore reserved for the most salient entities in the discourse model. The difference between demonstratives and pronouns, according to Gundel et al, is that demonstratives indicate that their referent is salient (activated), but that it is not the current most salient entity (in focus). Pronouns, on the other hand, can only be used for the most salient entities. In the literature, it is generally claimed that discourse-deictic reference, as opposed to individual anaphoric reference, is preferrably established with demonstratives rather than pronouns (Webber 1991; Asher 1993; Dahl & Hellman 1995). The contrast in (18) reflects these preferences: (18) [Jane bought [a new bike],]y. (a) It,'s great. (b) That's great. In contexts like this, where the predicate is great can conceivably be associated with either the referent of a full clause or an NP, the pronoun preferentially picks out an NP antecedent (a new bike), whereas the demonstrative picks out the whole clause [Jane bought a new bike). However, contexts that force either an individual or a discourse-deictic interpretation make it clear that both demonstratives and pronouns can be used for each type of reference: (19) A: I'm going to eat [the last piece of cake],. B: But John wanted to eat it,/that,. (20) A: I wonder whether I should [call him],. B: I wouldn't do that,/it, if I were you. In example 19, the anaphors occur in the object position of the verb eat, and must be interpreted as specifying a concrete entity. In example 20, the anaphors occur in the object position of the verb do and must thus specify Miriam Eckert and Michael Strube 61 an event concept. In spite of the preferences associated with the different NP forms, in each example both NP forms are capable of making the necessary specification. The observation that demonstratives are preferred for discourse-deictic reference is in line with the referent coercion assumption, i.e. the assumption that discourse-deictic anaphoric reference leads to the introduction of a new entity into the discourse model. If one assumes, following Gundel et al., that demonstratives are used for entities that are less salient than those specified by pronouns, then it is to be expected that demonstratives should be pereferred for entities newly created in the discourse model. 2.5 Right Frontier Rule We now move on to examining the structural constraints to which discourse deixis is subject. Webber (1991) notes that only text sections which are on the right frontier of the discourse structure tree are available for discourse-deictic reference, as can be seen by the following discourse (Webber 1991: ex. 14): (21) There's two houses you might be interested in. (a) House A is in Palo Alto. It's got 3 bedrooms and 2 baths, and was built in 1950. It's on a quarter acre, with a lovely garden, and the owner is asking $425K. But that's all I know about it. (b) House B is in Portola Valley. It's got 3 bedrooms, 4 baths and a kidney-shaped pool, and was also built in 1950. It's on 4 acres of steep wooded slope, with a view of the mountains. The owner is asking $6ooK. I heard all this from a real-estate friend of mine. (c) Is that enough information for you to decide which to look at? (c') *But that's all I know about House A. The central part of the text is clearly divided into two sections (a and b), each containing the description of a house consisting of more than one clause. At the end of each section a demonstrative is used to refer to what is described by the preceding utterances (that for House A; this for House B). Finally, in the continuation (c) the demonstrative that picks out the referents of the whole preceding discourse, i.e. what is referred to by (21a) and (b) together. The unacceptability of the utterance in the alternative continuation (c') shows that once section (a) is closed off and the description in 1 Although some NPs can function as antecedents to pronouns in the object position of do (e.g. do it/the foxtrot, do drugs), there is no number and gender compatible antecedent in the preceding clause in example 20. 62 Dialogue Acts, Synchronizing Units, and Anaphora Resolution {Info on both houses} 'But that's all I know . {info on House A} {info on House B} Figure I Discourse tree structure (Webber 1991) section (b) has started, (a) is no longer accessible for reference. Webber represents this discourse with the tree structure shown in Figure 1. The only nodes that a new constituent could attach to are nodes on the rightfrontier of the tree, which are indicated in the figure by the crossed circles. Asher's Principle of Availability (Asher 1993: 313) has a similar function to the Right Frontier Rule. It states in part that only the current constituent itself and its discourse referents and subconstituents (subDRSs) are available as antecedents for abstract object anaphora. Both Webber's and Asher's findings can be interpreted as reflecting the notion of adjacency. The constituents which act as antecedents to discourse-deictic anaphors must be linearly or hierarchically adjacent to their anaphors. We will make use of this rule in our algorithm, by formulating a concept of adjacency in terms of dialogue acts.