Centering Walker, Joshi, theory (Joshi Prince and 1998b) Weinstein is 1981; Grosz, of Grosz Joshi, and Sidner’s Weinstein overall 1983, 1995; and the component and theory ∗ Department of Computer Science, Wivenhoe Park, Colchester CO4 35Q, U.K. E-mail: poesio@essex.ac.uk. † Department Department of of Psychology, Computer Science, University University of Durham, E-mail: of Illinois U.K. ‡ at Chicago, Chicago, IL 60607-7053, USA. § MITRE bdieugen@cs.uic.edu Corporation, 202 Burlington Road, Bedford, MA 01730-1428, USA. E-mail:hitze@mitre.org. Submission received: 11 December 16 April 2003 2002; Revised submission received: 3 September 2003; Accepted for publication: © c 2004 Association for Computational Linguistics Computational Linguistics Volume 30, Number 3 of attention 1986) and coherence with in discourse (Grosz 1977; Sidner 1979; concerned salience, that is, Grosz and Sidner within local coherence A and of coherence is and it salience a discourse segment. fundamental characteristic centering viewed theory than computational one. By this we that mean is bet- ter as is a linguistic make cross-linguistically a which that its primary aim to process, from valid claims about discourses are easier or to abstracting (although away many specific algorithms for anaphora is algorithms are based on resolution anaphora The generation such In very different theory from those one usually finds in the theory). result a central Grosz, Joshi, Weinstein computational linguistics. provided papers such notions as and (1995), no algo- rithms are to compute such as working and “realization” that play a crucial role “utterance,” in the theory. “previous The utterance,” “ranking,” on centering argue while these concepts play a central role in researchers of that is any theory discourse coherence indeed, and salience, in way that some of their precise characterization (Walker, these concepts Iida, (e.g., Cote ranking) might best left for sub- sequent research, 1994). In other be defined a different notions for each language these inspired be viewed as parameters of and centering. This feature of words, should a great of the theory has (Kameyama deal research 1985; attempting Walker, Iida, to specify Cote centering’s 1994; Di parameters Eugenio 1998; for different Turan 1998; languages Strube and Hahn 1999). Competing and and of the theory have also been proposed: versions For of the central definitions of claims Gordon, center (CB) Gillion can (1993). be found in Grosz, example, Joshi, different and Weinstein definitions 1995) backward-looking Grosz, As a a researcher wishing (1983, and the of and centering, or it result, for practical applications, is confronted to with test predictions a The number main of to use large possible of measuring work instantiations impact in of the theory. goal the reported this of article was ways to explore of the space of parameter way of configurations, on the main the This different setting centering theory’s claims. identifying required specifying in the pa- rameters what an centering’s clearly the parameters, not explicit which claims are; previously discussed in the literature; and developing appropriate all of methods have instantiations (and been was statistical tests) out to carry out this evaluation. of English The comparison information by annotating a corpus texts from between with carried needed of instantiations different genres the to which test a variety centering corpus to the theory’s claims verified once and using this to in assess the extent way. The proponents of are the parameters are of set a certain is identify make centering have clearly stated that the clearly, aim the theory way to preferences is that cess; best to test such preferences through discourses easier to pro- and many the 1986; of Gordon, the Grosz, have in behavioral Dell theory fact Gillion been 1993; tested Brennan this way 1995). (Hudson, experiments, aspects But Tanen- haus, and number of made ways of and given the enormous possible only setting the means. theory’s A parameters, a systematic comparison other can be has well, by computational identify as among need which is that it is corpus-based evaluation advantages, aspects of perhaps the best way to the the theory or that to be further may specified, interact with and the prefer- factors such as temporal coherence ences (Also, stylistic variation that expressed knowing the to which the by centering. is important in its extent own real texts conform to centering Byron In preferences an Stent 1998; Eugenio studies 1998; of goal previous Di centering Kameyama (Walker right.) corpus-based 1998; Strube 1989; Passonneau Hahn 1999; 1993; Tetreault and 2001), only a instantiations of centering were and is more more few instantiations, in that it considers including a greater “crossing” number compared. instantiations of The present study well systematic parameters, in which as as parameter the parameters are set according to proposals due to different researchers. Only rePoesio et al. Centering: A Parametric Theory liable annotation techniques our were used; we other produced an well annotation manual can extend analysis to data, companion Web that be used to (http://cswww.essex.ac.uk/staff/poesio/cbc/) as as a not (The Web to allow readers to try out site tions in instantia- discussed this article. with site also of contains the a report full discussion all results.) Last annotation but not manual and technical not is more neutral a than in most previous studies in least, that, first of our evaluation we arguably proposing instantiations new instantiation of the theory; and secondly, all, are were a on all parameter The article is tested organized the same data. on as which follows. We we first (Constraint review the 1, basic Rule concepts 1, of Rule the 2) theory, discussing the three in claims We focus our and used formulation. then discuss how was and the parameters their was of corpus and the three main annotated Section how 4 the we annotation our main used to compute which violations in Section 5. claims. In present results, are discussed

It is not possible in to discuss in this of article work the in entire centering literature; we merely summarize this section in some of this For enough more detail to we allow the reader to follow the discussion the Grosz, rest Johsi, the article. Weinstein (1995) details, Walker, refer the Joshi, reader to Prince classic (1998b) references or such as of and in Poesio Stevenson and (forthcoming). and the discussion centering and 2.1 Centering Motivations As is and Main Intuitions simultaneously of it a theory of discourse coherence and of discourse salience . a theory coherence, attempts to characterize of way entity-coherent discourses: courses coherent because the discourse entities intro- dis- that are considered duced discussed. 1 At the same time, centering is also intended are and is, it which will most to be a theory of salience: that attempts to predict entities be salient at any given time. in which The main claim about local coherence mentioning made in centering is that discourse segments successive utterances keep in which the same discourse mentioned. entities are “more coherent” than Kintsch was discourse segments Dijk (1978) by Chafe Givon (1976) different (1983). and is entities are This hypothesis formulated In backed by empirical evidence is such as and van proposing and centering this hypothesis with further strenghtened by CB. Having that every utterance CB, it has is a unique “main link” the previous utterance: (Joshi Kuhn of the 1979; the inferences a unique required to integrate claimed, considerably Joshi Weinstein an utterance into simplifies the complexity 1981). the discourse Centering’s and first and contention as far as in an (more local course by utterance on salience is concerned is that realized realization below) are ranked: the dis- entities work on are more (Sidner salient 1979; than Prince others. 1981; This that is, that well, each is utterance some of much discourse entities Givon claim, as 1983; Gundel, a basic (Hudson, Hedberg, tenet Tanenhaus, Zacharski Dell 1993) discourse and 1986; and Gernsbacher is supported by much Hargreaves psychologi- cal Gordon, evidence Grosz, Gillion 1993; Stevenson, and Crawley, Kleinman and 1994). 1988; and and [BAR] Computational Linguistics Volume 30, Number 3 These claims identity about of coherence CB is and salience are linked by two further hypotheses: that CB the is most the crucially determined or is realized as a pronoun. by This the entities’ ranking and that the likely to be interpretation assumption of pronouns that is a “main entity” or “topic” in “focus” the preferred (e.g., Sanford Garrod 1981), commonly found in theories 1979) the psychological (Gundel, Hedberg, and Zacharski 1993) computational is (Sidner and linguistic and by such the between examples (1) literature and (2): and motivated evidence as contrast (1) a. b. c. d. (2) a. b. c. d. Something He must be wrong with has quite odd. John. been acting (He = He Mike yesterday. John) John called wanted up to meet him quite urgently. Something He must be wrong with has quite odd. John. been acting (He = He Mike yesterday. John) He called wanted up to meet him quite urgently. Grosz Discourses (1d) (1) is and not (2) differ only only (2d). in The their (d) sentences, but (c) felicitous as reason, they argue, is according to et al., as next provided is it more is in it; than Mike, that of discourse entity so after utterances, John highly ranked it will the the CB the be pronominalizing the utterance CB, that pronominalized realized and if given the preference for This the pronominalization John should be identity of anything Grosz the the CB else is. link between and has been used by CB (contra, et al. to Sidner support [1979] the claim whose discussed above that utterances Grosz have a unique e.g., (3c)–(3f) theory of assumed two foci). initiated et al. note the (3a)–(3b): contrast between continuations the discourse by utterances (3) a. b. c. d. e. f. Susan She She gave Betsy a pet hamster. reminded Betsy her whether that such hamsters were quite shy. Betsy asked she liked the gift. Susan told her that whether she really liked the gift. She asked Susan her she liked the gift. told that she really liked the gift. Grosz et al. argue were that continuations (3c)–(3f) (3b), are less and less acceptable, whereas if Susan and Betsy equally ranked after all variants should be equally accept- able. 2.2 Terminology and Definitions 2.2.1 Local Focus, Forward-Looking is processing Centers, and Utterances. involves A fundamental assump- tion underlying centering that or a . The discourse includes continuous of updates to the local attentional (CFs), state, which local focus Sidner’s local (1979) focus a set forward- looking centers mentions correspond of to (Karttunen “potential 1982; viewed as Reyle 1993). The discourse entities information 1976; discourse Webber 1978; foci” and Heim can be prominence Kamp and In or of these CFs CFs. local The focus also contains about the relative : rank local focus new gets ones, updated after terance the replaced by and the CB every ut- this update, current are changes, Poesio et al. Centering: A Parametric Theory as well (see below). 2 The (DS) set is of CFs indicated introduced CF(U in the ,DS), local focus by utterance U a segment by i generally abbreviated to CF(U i in discourse Brennan, Friedman, means of Pollard one of (1987) i ). CFs and formalized the relationship between utterances and by their so-called constraints: 3 Constraint 2: Every element of CF(U,DS) must be realized in U. 2.2.2 Ranking, Preferred important Centers, of and Backward-Looking Centers. We already men- tioned two of claims the theory: CFs that forward-looking prominence. centers this acquire particular The are ranked, and that because is only ranking, some (when one to be partial, but the most highly ranked CF ranking function required is (CP). Ranking realized called Preferred Center is by an utterance to one of exists) the CFs the the CB. The CB is the closest in also used characterize notion of as (Sgall 1967; Chafe 1976; Sanford concept Garrod centering to Givon the traditional 1983; Vallduvi 1990; “topic” Gundel, Hedberg, Zacharski 1993) and 1981; role in and Although and plays Joshi, theory’s claims about both coherence and salience. in Grosz, a cen- tral and Weinstein the work (1983), the CB on was characterized only in intuitive (Grosz, Joshi, terms, most subsequent 1995), has been based 3” the Brennan, definition Friedman, below Pollard and Weinstein referred to as “Constraint by and (1987): Constraint 3 CB(U i ), the of backward-looking CF(U ) is center of in utterance U . U i , is the highest- ranked element i − 1 that realized i Notice on that according to this definition, making the computation of the CB depends important exclusively ranking and “previous utterance,” these parameters crucially for the framework. 4 2.2.3 Transitions. The hypothesis that discourses are easier to process perceived “about” discourse is when successive utterances in are of as being induce classification of a unique entity formalized terms utterances Many according to the type of in centering (update) a proposed; they in local focus. such classifications of transition Grosz, Joshi, the Weinstein (1995) transitions have been on whether and distinguish among of U three transitions, the backward-looking i 1 is types maintained of or not in U depending i and on whether CB(U i ) is most center − also the highly ranked entity (CP) of U i : Center Continuation CF (CP) (CON): of U (i.e., CB(U CP(U i ) = CB(U ) CB(U i − 1 ), )) and CB(U i ) is the most highly ranked i i = i Center Retaining (RET): CB(U i ) = CB(U i 1 ), but CP(U i ) = − CB(U i ) Center Shifting (SHIFT): CB(U i 1 ) = − CB(U i ) We will consider a few alternative classification one of schemes below, after of discussing are formulate the core claims centering, Rule how these classifications used to 2. Computational Linguistics Volume 30, Number 3 2.3 In Main Claims the words of Grosz et al., the most fundamental claim its of centering is will that “to extent adheres coherence increase the that discourse inference to centering constraints, 1995, the hearer will decrease” (Grosz, Joshi, and the load placed 210). upon page Grosz list seven such “costraints,” three of and which Wein- stein Even et al. we not can be directly evaluated. though introduced are in following Brennan, here Friedman, the distinction Pollard between “constraints” will and “rules” now for these three claims the names Brennan and (1987), et al. gave them, by which we use they are best known: Constraint 1 one (Strong): CB. All utterances of a segment except for the first have exactly Rule 1 (GJW95): If any CF is pronominalized, the CB is. Rule 2 (GJW95): (Sequences which of) continuations over (sequences are preferred of) over (sequences of) retains, are preferred shifts. 2.3.1 Constraint 1, of Topic Uniqueness, idea of and Entity (Vallduvi Coherence. 1990; Gundel If we 1998; Hurewitz view the CB 1988; as a Miltsakaki formalization from 1999; Beaver the Kuhn 2004), (1979) Constraint “topic” Joshi Joshi 1 expresses, Weinstein first (1981) and foremost, the original claim one (or no and more one) and and that (1979) at each point are easier may to process. discourses This with exactly with Sidner’s than “topic” view contrasts with both Givon (1983), hypothesis Alshawi that (1987), utterances Lappin have Leass two (1998), such as matter of and (1994) “topics” and Arnold theories number which and of view “topichood” as a degree and therefore allow for an arbitrary In It the strong topics. form just presented, Constraint 1 is also a claim about local coher- ence. expresses a preference Each for discourses in to be entity coherent: to continue in should A weaker realize at of least Constraint one talking about the same entities. utterance a segment of the discourse entities realized (e.g., the Walker, previous Joshi, utterance. Prince 1998a, form 2, 3); 1 has also been suggested CB is and not footnote page the preference for a unique preserved, but the preference for entity coherence. Constraint 1 CB. (Weak): All utterances of a segment except for the first have at most one 2.3.2 pronominalization. Rule 1 and Pronominalization. In Rule 1 is it the main claim of centering nalizing presented above, states a preference for pronomi- about CB, The if the version the original anything is pronominalized of in at Grosz, all. We Joshi, also examined Weinstein two alternative formulations. form the claim and (1983) was as follows: Rule 1 (GJW83): If the CB of pronoun the current utterance is the same as the CB of the previous utterance, a should be used. Gordon, Grosz, and Gillion in (1993) proposed a much in stronger form were of the pronominalized ( certain positions the sentence read more claim. They found that entities realized [RNP]). This slowly unless repeated-name penalty 5 evidence led them to Poesio et al. Centering: A Parametric Theory propose a more RNP—for restrictive definition of CB (briefly, that of the CB is the entity subject see longer version this article available on to the discussion, stronger of the Rule 1, CB (defined in more the Web site) and way) a form pronominalized: requiring the this restrictive to be always Rule 1 (Gordon et al.): The CB should be pronominalized. Although in we mind will refer to this version as “Gordon of CB proposed et al.’s” for brevity, readers keep that because the Gordon should of Rule definition 1 is properly only by et al. is more restrictive, (The their with version instantiation evaluated on are discussed in using that the longer version of definition. results Web this the article availabel the site.) 2.3.3 well: Rule It 2 and the Classification of Transitions. CB over Rule 2 is a a preserving the it claim about coherence, preference changing and for realizing it as most states of entity over for its relative of ranking. This aspect of as the salient changing the a attention; several variants this constraint have been proposed, theory has received well many lot The ways of classifying transitions. We studied many alternative proposals. as as version of Rule 2 presented of in Grosz, (e.g., Joshi, CON-CON and Weinstein over SHIFT-SHIFT) (1995) expresses preferences among sequences transitions Di Eugenio (1998), This form of the constraint is in rather than preferences for particular transitions. part mo- tivated by empirical of null results. pronouns in Italian for example, Continuations explicit CON depends or SHIFT, on found that the relative distribution well: in Center and the it is previous null follow a a much more transition as a pronoun will that used, whereas in Center Continuations likely that null be pronouns likely. Turan that (1998) follow obtained a RET transition, both null and explicit pronouns are in equally similar Other results for and explicit (Brennan, Friedman, argue instead Pollard the inferential Turkish. researchers that 1987; Walker, load Iida, is evaluated Cote 1994; utterance Walker, by utterance Joshi, Prince 1998a). The and of Rule 2 proposed Brennan and and version by et al. is as follows: Rule 2 (Single Transitions): RET Transition states which are is ordered. The CON Smooth transition Shift is preferred to which the is transition, Rough preferred Shift to the transition, preferred to the transition. This SHIFT: formulation Smooth Shift of Rule (SSH), 2 depends when CB(U on a further ) CP(U distinction ), Rough-Shift between two (RSH), types when of ) ). Transitions n = CB(U CP(U n and n = n can then be classified along two dimensions, as in the following table: CB(U n ) = CB(U n 1 ) or CB(U n 1 ) = NIL CB(U n ) = − − CB(U n − 1 ) CB(U n ) = CP(U n ) Continue Smooth Shift CB(U n ) = CP(U n ) Retain Rough Shift (1986) Further proposed refinements of these classification Center schemes Establishment have been proposed. (EST), Kameyama a fourth transition type, for utterances E.g., in Bruno more was the when bully of Bruno the neighborhood. was Bruno when / He often He taunted was Tommy , the second sentence would be read slowly used than used. Computational Linguistics Volume 30, Number 3 that establish ment. Walker, a CB Iida, after an utterance Cote (1994) without one, such as the first utterance of a seg- Center and Continuations, idea argued that these utterances first should of be classified as ond CB, but this is CB is the initially being that even the underspecified Notice and is utterance determined strong only a segment processed. when does have a of the 1, according to the version Constraint sec- utterance the CB in first of that utterance a discourse segment none of is the only utterance allowed not to have a includes a coherent discourse; hence, inverse of these Center classification Establishment, schemes (ZERO) the CB one that which we for might transitions roing classes either for CB-less (a call Ze- transition less other utterance CB-less following ones (the NULL does have a CB) or for Strube utterances Hahn (1999), following across Grosz, Joshi, and Weinstein (1995), transition). and is like (pairs, in of claim that infer- ential load way evaluated of sequences inferential of fact) transitions In but a the load utterances. their view, argue for fications different of evaluating transitions do not reflect what should be one classi- crucial In claims of such as those the theory: that the CP above of one utterance predicts the CB of of the next. the on order to formalize their view, they propose a different classification 332): between cheap and expensive transitions (Strube scheme, and Hahn based the distinction 1999, page A transition pair CP is cheap of if the CB of the current utterance CP(U the the previous utterance, that is, if CB(U is correctly predicted ). by n )= n − 1 A transition pair is expensive if CB(U n ) = CP(U n − 1 ). Strube and Hahn then propose a new version of Rule 2 based on this distinction: Rule 2 (Strube and Hahn): Cheap transition pairs are preferred to expensive ones. 6 2.4 Although The Parameters Grosz of Centering et al. discussed possible definitions for the concepts used in the claims above–“utterance,” “pronominalization” on “previous utterance,” Rule English. “ranking,” 1. But without Similarly and “realization”—they settle for undefined is the notion didn’t a specific definition, even impossible by further specification of of it is governed to the claims above, just it is not these concepts of, say, evaluate as without possible the “government binding theory” providing to evaluate predictions of of “command” with or and an explicit definition “argument”. As a result, a considerable what amount essentially, research has been concerned of theory. We establishing the best specification of proposals for in are, parameters the briefly review some these this section. 7 [BAR] Poesio et al. Centering: A Parametric Theory 2.4.1 Utterance implicitly and Previous identified with Utterance. In the early Kameyama centering (1998), papers, however, utterances were identification makes sentences. number of of argued that such much if the were potential antecedents anaphoric expressions greater than they identification resolved problems clause with by clause. multiclausal Furthermore, she noted that this leads to more one becomes Kameyama difficult sentences: for example, grammatical function ranking proposed to compute, as a sentence may have than subject. not that the local focus be updated into (1) after every tensed clause, after every sentence, and of classified tensed clauses utterances that constitute a “permanent” (2) update the local focus, in such as coordinated clauses and adjuncts, and much embedded information utterances introduced that result into temporary updates that are then “popped”, is as the Grosz Sidner discourse (1986). by According subordinated discourse Kameyama, segments only popped a of according to and types such as the complements of to are For few example, Kameyama clauses, proposes to break up (4) into certain verbs, embedded. of including utterances as follows, (U2) or (U5), and to treat each these utterances, subordinate clauses such as as an update: (4) (u1) (u2) Her entrance in Scene 2 Act 1 brought even she had sung note. (u3) some Thereafter disconcerting applause waxed before (u4) a operagoers the audience applause (u5) happy but discriminating of reserved judgment as her singing showed signs strain. Experiments in main by Pearson, Stevenson, more and Poesio (2000) confirmed mentioned that CFs introduced introduced clauses in are significantly However, likely to semicontrolled be subsequently Suri than of a study by and Mc- CFs Coy (1994) complement other clauses. suggested that types clauses—specifically, by and before–are also “embedded,” not adjunct clauses headed Kameyama; after were “permanent updates” Cooreman as suggested Sanford by (1996). The these of results other subsequently of is confirmed clear. Kameyama by (1998) and status of types clauses less tentative relative clauses, according which also pro- posed a analysis to and the local focus then merged they with are temporarily treated as utterances update clause; of didn’t, however provide but are the embed- ding she empirical support not for this in hypothesis. Other types Strube subordinate (1998) clauses Miltsakaki and parentheticals with and Miltsakaki (1999) are ances (1999) question Kameyama’s discussed argues, on identification this literature. (tensed) the basis of data from of English utter- CFs Greek, clauses. and in main that the local focus is updated when only after every sentence CB. and that only the the clause are considered establishing the 2.4.2 Realization. U to Grosz, Joshi, C and Weinstein on (1995) simply say that what They which depends the particular semantic theory one it means for utterance realize ways center in a entity may be “realized” in adopts. consider two Constraint 2. discourse is when noun in an utterance as required by entity. Direct realization is when a one phrase of noun the utterance refers in to that discourse Hawkins (1978), is Indirect realization the utterance an associative reference to that discourse entity in phrases the is, object which the sense mentioned anaphoric expression that refers to an wasn’t of 8 that in but is an before somehow related to an object that already has. For example, the following discourse: (5) (u1) John walked toward the house. (u2) The door was open. Computational Linguistics Volume 30, Number 3 John, thehouse and thedoor are directly realized indirectly in the respective in (u2) utterances; can being realized by virtue of in addition, the house be thought as to reference the door (see, the discussion in being the [1995] Joshi, e.g., Prince [1998b]). Clearly, Grosz, referred by Weinstein associative Walker, Joshi, and CB is and (5), which and the entities are considered to be “realized” in computation of the affected by in (u2) example, (u2) of has its a CB (the house) only with if the house To is an utterance: In for by associated the door. our considered to be realized virtue being notions of on of knowledge, the effect these predictions theory not of alternative realization the of the (discourse) in focusing in such as Sidner’s has (1979) been previously studied, even though theories issue do be these and the is allow the mentioned in focus to of realized an utterance cases, often A discussions of issue entity. is whether centering. related Many empty of realizations, or traces , should morphologically count as realiza- tions an occur in theories grammar hypothesize that control structure underlying a variety of null elements ing the syntactic (VP) in (6a), reduced relatives as in (6b), constructions, includ- constructions in (6c): as and even coordinated verb phrases as (6) a. b. c. John John John wanted ( ∅ to buy bought a house ( a house). ∅ abandoned ( promptly by its previous a and ∅ demolished it). occupiers). bought house If, or for not example, it the coordinated of VP in (6c) is is considered a separate CB. realization John going determine whether utterance, contains to or not whether To our a morphologically other null English. have been considered in it has a knowledge, only elements the centering Eugenio An issue literature for languages 1998; that Byron has been Stent raised 1998) in than the is whether centering literature CF (e.g., Walker only 1993; Di and noun (NPs), or the list contains as the entities realized as first- entities realized third-person of however, NPs. Walker if in (1993) phrases also and second-person suggests pronoun (7), neither that deictic entities is utterance not (u2) nor are beyond the introducing utterance (u3) purview centering; CB example would have CF a second-person you counted as an entity in the list. 9 (7) (u1) (u2) You if should not use PRODUCT-Z (u3) Whilst you are pregnant of breast-feeding. you are receiving PRODUCT-Z ... 2.4.3 Ranking. of Perhaps the most discussed parameter of CB of centering, Constraint at least in versions accept the definition specified by 3, the the theory Most that working on including Grosz is the ranking function. researchers role in centering, several play determining the relative ranking of et al., assume that centers; in factors Walker, a Iida, and Cote (1994) and Walker, Joshi, forward- looking (1998a) fact, may not in and Prince claim that the factors affecting ranking be the same all languages. [BAR] Poesio et al. Centering: A Parametric Theory ≺ 2.4.4 R1-Pronouns. not Rule 1 states that which if any CF of is “pronouns” pronominalized, the CB is, but the theory It does explicitly specify types is, if clear that realization as a third-person singular pronoun are covered by this rule. seems is pronoun does count; that the choice between using a third-person singular to realize a [BAR] Computational Linguistics Volume 30, Number 3 CB or another Italian, CF, Japanese, the CB should Turkish, be chosen. We also saw of that CBs in languages such as null and But (Kameyama 1986; the preferred in Walker, English Iida, realizations Cote 1994; Turan are 1998; morphologically elements if Di CF Eugenio pronoun, verifying the rule a is 1998). should an utterance is the CB count as with as a trace? Or if the CB is realized realized with as a full NP, a third-person but a second CF and realized pronouns? The a demonstrative pronoun? of And what about theory, 1, which we will precise characterization is the (sub)class of first- and second-person Rule pronouns subject to of our call R1-pronoun no proposals s, clearly yet the best knowledge, in an essential aspect of the to this regard can be found in the centering literature. 2.5 Empirical Centering Support for, and Applications of, Centering natural has served processing as the (NLP), theoretical foundation psychology. for This a lot of work ies testing claims of and includes in linguistics, language the theory for languages including English, annotation German, Hindi, stud- Italian, Japanese, the Cote 1994; Di and Turkish Eugenio (e.g., 1998; Kameyama Turan 1998) 1985; Passonneau 1993; in Walker, (Walker, Iida, and Prince (1998b). The pronominalization and several made papers in Joshi, and claims about (Brennan, centering to both for anaphora resolution Friedman, have been applied Pollard 1987; develop Strube algorithms Hahn 1999; Tetreault 2001) (Dale and 1992; Henschel, of Cheng, and Rule Poesio 1. Ideas 2000); from this work and for sentence planning and can be viewed in as providing in centering, 1985; and Kibble particular Power Rule 2, an evaluation increasingly claims such as 2001; useful text planning (McKeown and 2000; are Knott found with We et al. Karamanis 2003). have already seen that In some many predictions of of the theory have also been pronominal in in these experiments, differences in process- tested ing psychological techniques. (according references instantiation a sentence of a sentence were to entities observed: with observed the theory) introduced in Hudson, different ranks to a particular pronominal for ex- ample, in that interpreted references more to entities nonpronominal subject position or the previous sentence nonsubjects are (Hudson, Tanenhaus, quickly than Dell 1986; Hudson-D’Zmura references references and (1993) Tanenhaus to identified 1998). processing And we already mentioned and RNP, that when Gordon, NPs Grosz, a slowdown, the in and Gillion introduced time in or first-mentioned subject in position referring to entities not pronominalized. subject position the previous sentence However, are the discussion in in this section many should have ways made it clear just our no has and how different they in can be instantiated. how many parameters To the theory knowledge, instantiation previous of study more has attempted one of to analyze a systematic way how of varying theory, the than original This for combinations is of of these parameters parameter work settings not affects the claims the especially considered in the papers. analysis the goal the discussed here.

Given way the make many ways in which the parameters of centering can be set, the only feasible to a means. systematic That is, comparison between the theory’s of “instantiations” process of is by computational running computer simulations the using corpus and comparing the results obtained local focus update of instantiations. an annotated The evaluation principle we under different number of use when for this comparison is the in “violations” way (e.g., whether the pronominalization theory’s claims resulting in the parameters with are Rule set a certain choices are accordance Poesio et al. Centering: A Parametric Theory 1). we In this section our we discuss how we methods, set about performing such was comparison, the data used, annotation and how the annotation used. 3.1 A Evaluating the Claims of we Centering against preliminary is what a Corpus theory. As question in had Section to 2, address of are in fact the mentioned main in claims Grosz, of Joshi, the Weinstein discussed (1995), Constraint 1, Rule the 1, seven claims Rule 2 ones and we and on are the Because that can actually verified concentrated these. several of be using a corpus; proposed, we of variants well. these three claims have been evaluated a few these variants as 11 The second important what we question is how these three claims are meant can corpus to tell about them. The proponents to be inter- preted and expect a clear the theory does not us of centering is, are quite of whose that state “hard” facts about language, that 1, the Rule kind 1, facts Rule 2 violation meant leads to ungrammaticality straint be preferences which, judgments. when Con- and process. are not, to 12 The to mere For one presence of followed, lead to texts that are easier we a few exceptions to a claim does therefore, interact count with as other a falsification. (a thing, not should expect centering preferences to And another, constraints may point no way emphasized of for there be of enough in the centering information literature). without expressing in such preferences. 13 So a particular piece violating some is, at best, expect to be verified statistical that that the number we can of the three claims utterances claims will a sense: that verify such in be we significantly may find higher than the number of utterances that violate will them—and not fact, As we that for some claims, even statistical significance Page we be achieved. a result, tests; are the test for Constraint evaluated 1 each and Rule claim statistical is Rule tests used sign 1, using the important 2 (Siegel and Castellan olations” to keep in mind 1988). and the It test for also in processing difficulties: that a corpus This cannot tell us whether only “vi- actually result So we can be determined minimizing as reading-time experiments. should make it by behavioral studies such clear should not the only absolutely that violations Nevertheless, cannot and be instantiations combinatorics in other of the way. problem deciding Furthermore, make factor it im- in theorizing about centering. the possible of to compare is the parameter most way any may also the systematic to identify other this form evaluation interact with We issues preferences in Section 5. and constraints that centering. return to these

The data used from in this work are texts The from museum the GNOME corpus, which museum objects three domains. subcorpus consists of currently in- cludes texts who produced descriptions and the them. 14 The of is brief texts about ceutical selection of artists leaflets providing patients with mandatory pharma- subcorpus a legally Computational Linguistics Volume 30, Number 3 information from about their medicines. 15 Sherlock The GNOME corpus University also of includes Pittsburgh tutorial (Di dia- logues Moore, the first Paolucci corpus 1997). Each collected at the subcorpus main contains about 6,000 NPs. Texts Eugenio, and were used the experiments reported here. The from the two domains corpus was for the segmentation experiments discussed in third sub- used for on Web the extended technical report available The the site. interest. data First used of for this study have not two characteristics that make in them of more cover genres previously considered studies on particular all, they to strongly that “real” NLP applications (see, Knott have to [2001] contend with. centering and similar those At the same of time, museum they are entity-centered e.g., et al in for an analysis the data), important so the hypotheses in way about coherence formulated centering are likely to play an part the these texts are constructed. 3.3 The Annotation 1989; previous Passonneau corpus-based 1993, investigations 1998; Byron of Stent centering 1998; Di theory we Eugenio 1998; are aware of (Walker Kameyama 1998; or Strube Hahn and 1999) were our corpus to her or all carried own out Hurewitz 1998; and by a single annotator annotating (Passonneau was her his according Litman for this study only information his subjective judgment. One of goals to use 1993; Carletta 1996), we that could will be annotated easier The we as believe this is make we our reliably and results proposals to replicate. price paid of to achieve replicability proposed that in couldn’t test all about the computation centering parameters ture, about segmentation and about ranking, discussed below. the litera- especially those manual, on as annotation detailed available the companion Web Eight The followed were a involved in site. paid annotators we information the reliability we studies were and the annotation. what In the following briefly discuss problems the we that more able to annotate, annotate, encountered; for details, we we didn’t and the to extended of the article and the Web refer readers the A version prohibitively systematic with among different ways site. comparison of setting methods, the parameters it’s not would be with expensive either. traditional to corpus or one Obviously, psychological way fixing it but easy do analysis, of can’t be done by directly annotating “utterances” in most “CB” according of to (Byron the parameters, Stent 1998; Di as has Eugenio been previous centering theory and 1998; done Kameyama with 1998; studies Passonneau 1993; Walker 1989). instantiations Instead, we of annotated theory, our corpus the primitive concepts used by one different or other instantiation the of its by notions. the This includes, of that is, in- formation role in that has been claimed centering to play a the into definitions basic for example, how sentences break up including clauses and subclausal units; An grammatical function; and information anaphoric re- lations, bridging CF references. automatic CB, script uses this utterances, ranking, and their according particular way of to compute their to a setting the instantiation. parameters, and to compute statistics relevant to the three claims according to that 3.3.1 Utterances. (sentences In order finite to evaluate the well definitions of utterance proposals proposed in the literature versus clauses), as we marked as the different of concerning the “previous utterance” discussed This above, includes all spans (defined text that might of be claimed to update the local focus. sentences as all units text Poesio et al. Centering: A Parametric Theory ending ( with a period, ) a question . Units mark, include or an exclamation (defined point) as well of as what we called discourse units its obligatory clauses as sequences text containing a verbal other complex, all arguments, might and all postverbal sentence that independently adjuncts) as well as subconstituents (PPs), update the (the local focus, such as parentheticals, VPs. preposed prepositional phrases and second element of) coordinated 16 Sentences interrogative, have one attribute, imperative, or stype , specifying The whether of the sentence include is declarative, exclamative. attributes units • utype : whether the unit is a main clause, a relative clause, an appositive, a parenthetical, etc. • verbed : whether the unit contains a verb or not. • finite : for verbed units, whether the verb is finite or not. Marking up sentences proved to be quite The easy; marking on up units, on of training. agreement units, kappa discussed in Carletta identifying the other hand, required extensive annotator using statistic (1996), was the boundaries the κ = 500 on (two . 9 (for 200 two annotators was and units); the : agreement 76; features as utype : 9; annotators finite : 81 and units) 505 verbed κ = κ = . . . In at least follows: κ = main . more 1,000 total, finite study contain sentences and than units, including the texts used 900 for the clauses 3.3.2 in NPs. MATE Our instructions for identifying NP markables (Poesio, derive mary for annotating anaphoric relations Bruneseaux, from those proposed the 1999), in scheme from DRAMA (Passonneau 1997) MUC-7 (Chinchor and Ro- Sundheim 1995). turn derived In In GNOME NPs marked and and the corpus, are 3,345 by NPs. < ne which texts used 217 for this study contain These > (‘Nominal pronouns, include en- tity’) 586 pronouns, elements. total, the 23 among 250 308 are third-person second-person 391 proper pronouns; personal nouns; 1,290 and possessive the-NPs, demonstratives, and NPs, 1,119 definite indefinite NPs, NPs, including including 554 745 NPs possessive 114 bare and NPs. 269 a-NPs; and and 350 other NPs, including 117 quantified NPs and coordinated We annotated properties 14 attributes (Poesio of 2000). < ne > Those elements specifying their syntactic, semantic, and include discourse relevant to the study discussed here the following: • The np type, cat , with values such as a-np , that-np , the-np , and pers-pro . • The which agreement features of num pronoun , per , and gen , used identified to identify unambiguously. contexts in the antecedent a could be • The from grammatical in function FRAMENET gf . Our instructions project (Baker, for this Fillmore, feature are 1998); used the and Lowe derived those , see , also http://www.icsi.berkeley.edu/ (used objects in ˜ framenet/. The values are subj obj predicate for postverbal ), (for copular objects sentences, there-sentences), watch) (for indirect there-obj objects), postverbal (for in such as This is (a production PPs modifying VPs), comp (for in adjunct the gen s determiner position in argument of NPs), , , np np-part , adj-mod , and no-gf (for possessive np-compl in np-mod np s occurring by themselves—eg., titles). Computational Linguistics Volume 30, Number 3 The .84; agreement : .9. We values for these attributes problems are as follows: with cat: .9; gen: .89; gf : information .85; num : per number encountered gender, even with supposedly “easy” as so semantic attributes (see such manual and on but especially Web in We were, however, One able to mark the annotation available the site). is up relevant this fashion. exception we the attributes for study a reliable to acceptable (Sidner on Cote feature 1988; a of NPs Stevenson, often that weren’t able reach roles agreement 1979; claimed Crawley, to affect Kleinman ranking, thematic in was 35. As we and were not 1994); the agreement value this case κ = on roles. . a result, able to evaluate ranking functions based thematic (8) <unit finite=‘finite-yes’ id=‘u227’>

<ne id=‘ne546’ gf=‘subj’> The drawing of

<ne id=‘ne547’ gf=‘np-compl’>the corner cupboard </ne></ne>

<unit finite=‘no-finite’ id=‘u228’>, or more probably

<ne id=‘ne548’ gf=‘no-gf’> an engraving of

<ne id=‘ne549’ gf=‘np-compl’> it </ne></ne> Poesio et al. Centering: A Parametric Theory </unit>, ... </unit> <ante current="ne549" rel="ident"> <anchor ID="ne547"> </ante> Work nary such as Sidner (1979) and Strube indirect and Hahn (1999) as well as our However, that realization can play a crucial role own in prelimi- CB. analysis, suggested in attempts at marking anaphoric information, maintaining the of previous MUC identity initiative, suggested that while it’s particularly the context on the marking fairly is easy to achieve agreement was relations, Poesio up Vieira associative work, we as marked only and (1998). relations of For quite hard; this confirmed by studies such these reasons, Besides a few types relations, and we and to reduce the annotators’ membership identity ), ( IDENT ( ), we marked ), up only specified priorities. ( three associative ( relations: subset and “generalized possession” ), which set includes ELEMENT well SUBSET ownership We marked only POSS objects part-of relations as noun as not, relations. relations between realized or propositions by phrases implicitly and introduced for example, anaphoric or references to actions, events, instructions our by clauses much sentences. gave to annotators concerning how to mark. (See We also strict the annotation For manual in available on the Web site.) , Furthermore, NP we ences: the Dauphin the embedding would specified prefer- example, Francois, NP in be chosen as an antecedent, As rather we than the appositive our most found a reasonable (if not position. In expected, (two marking perfect) agreement on identity relations. 200 NPs) we recent observed analysis no annotators 79.4% of the anaphoric relations one in the relations of were marked between by 12.8% real only disagreements; of them; and 7.7% the one of up both annotators; marked by other. With cases, the annotators up a closer antecedent than the 18 associative (only 4.8% references, limiting the relations did marked curtail the disagreements only among 22% of annotators of are differently), but bridging references were the relations marked one or in actually other the same annotator. way by So both annotators; 73.17% on of relations information were involved marked by only the reaching more agreement one this over (Poesio several discussions between annotators and than pass the corpus 2000). 3.3.4 Segmentation. only within According to Grosz and Sidner A proper (1986), centering ture discourse segments. evaluation of is meant to cap- preferences Unfortunately, would require a corpus in which the claims of the identified. theory therefore difficult discourse identify segments have (Passon- been neau Litman 1993), discourse Grosz segments Sidner are (1986) to and and do not provide reliably of of intentions and Grosz that it can Sidner, be used to identify a specification discourse explicit enough the intentional structure As texts—which, only according to and determines their a at annotating texts according to Grosz segmenta- tion. Sidner’s result, preliminary made. attempts and For theory have this most been reason, or previous corpus-based proposed studies of centering Walker (1989): either Consider ignored segmentation used heuristics such as those by every paragraph pronoun in as a separate discourse or segment, pronoun except whose when its first sentence con- tains a subject position a agreement features are not [BAR] Computational Linguistics Volume 30, Number 3 matched well, by any other CF in the same of sentence. We tested rough only using layout structure the texts indicator heuristic of methods as In the we only as a with discourse article the the heuristic proposed structure. this discuss results on Web we by Walker. In the extended other technical report available well the site, discuss with the results with segmentation of heuristics, GNOME as as further results independently the tutorial dialogues subdomain the ner’s proposals, (Moser corpus, from and Moore Grosz 1996), annotated according technique inspired to which a by Grosz relational discourse analysis Sidner–like was and Sid- proposed in Poesio Di Eugenio a and (2001). segmentation extracted as and 3.4 The Automatic Computation ing corpus is of Centering (utterances, used by Perl Information annotated CFs, scripts CBs) that automatically compute the center- instantiation data structures find of and Constraint according 1, Rule to 1, the particular Rule 2 (according parameter chosen, of Rule 1 violations Rule 2), and to several The versions of and is controlled and evaluate number the claims of using the including statistical tests. behavior the scripts by a parameters, CBdef:

which were definition of CB should be used. (All in the Constraint results discussed 3.) in this article computed using the definition

identify utterances with sentences, finite clauses, or verbed clauses. previous utterance: treat adjunct clauses Kameyama-style or Suri and McCoy- style. uttdef: realization: allow only direct realization, or allow indirect realization as well. CF-filter: ment treat it all is NPs as introducing omit first- CFs, or exclude certain NPs classes. At the NPs mo- possible to (e.g., and second-person in and/or predicative policeman John is policeman ). in position a a rank: rank CFs according in Gordon, to grammatical Grosz, function, Gillion (1993), linear or order, information a combination of the in Strube two as Hahn (1999). and status as and prodef: or consider include as R1-pronouns only pronouns third-person ( personal pronoun demonstrative that, these), pronouns ( it, they), also ( you). and/or the second- person segmentation: identify or whole segments using Walker’s heuristics, or with paragraphs, sections, texts. Among will the mention many other script parameters whether whose implicit effect will not be discussed in here refer- we just those that ences be treated CFs, determine the relative ranking of anaphors entities in bridging should as (See complex NPs, to clauses. the extended technical report on and how Web handle is The “preposed” adjunct used to compute statistics concerning violations of the site.) algorithm find straightforward, in and we will therefore omit The it one here; the interested the claims fairly reader can we need a discussion mention the extended is technical pronouns. report. to pronoun As it additional complication that here relative is controlled could be argued that the decision to we generate a relative not ignore as much primarily as possible, in by grammatical considerations, attempted to them (1995) if only as “pronoun” a violation/verification non-CB of the Rule following is 1 from sense. Grosz, Our Joshi, scripts do Weinstein count an utterance CB realizing a a relative pronoun, and is only the pronoun. What means in is or the realized by a relative this practice that the Poesio et al. Centering: A Parametric Theory Table 1 Number of utterances and CFs with the vanilla instantiation. Museum Pharmaceutical Total Number of utterances 430 577 1,007 Number of utterances that are segment boundaries 91 134 225 Number of CFs 1,731 1,308 3,039 number of of utterances with CB, examined we will to evaluate Rule shortly. 1 is generally less than the number utterances a as see

Given with the number of instantiations. parameters, Instead, it we is difficult, if not impossible, to instantiation,” begin by results with discuss what the all discussing the we results on most often call the in “vanilla of theory. based We the settings for the parameters of then examine the results obtained used discussions the we of After by varying the definitions utterance and realization. establishing the “best” Additional values for these parameters, are Readers who want in consider the effect alternative ranking functions. the extended out instantiations technical report available on the companion Web results discussed to not here should the Web site. try discussed visit site. 4.1 What The we Vanilla Instantiation call the “vanilla instantiation” is not an instantiation actually proposed of come as close as possible to a “mainstream” instanti- in the literature, but an attempt to ation Brennan, centering Friedman, by blending Pollard proposals (1987) from incorporating Grosz, Joshi, and Weinstein (1995) Kameyama and additional suggestions from and (1998) and on and Walker, of CB from Joshi, Grosz, and Joshi, Prince (1998a). Weinstein The vanilla matical definition and (1995) instantiation is based the (1987). function for ranking, as not proposed provide there and in of Brennan, Friedman, and uses incorporates Grosz Pollard gram- Because et al. do from Kameyama a definition (1998) utterance, the vanilla instan- tiation the hypothesis of proposed that utterances are and characterization there. 19 Concerning finite clauses in the instantiation “previous in only utterance” realization, the vanilla third-person only if NPs it is introduce mentioned. CFs, and a discourse For entity counts including only of Rule as “realized” (and 1, we mainly an utterance explicitly the purposes studied pronouns a “strict” definition (see possessive) and relative pronouns of R1-pronoun Walker, Joshi, personal Prince [1998a], 4), we and “broader” traces including and page pronouns but this, also include NP modify, that, considered these, a definition the demonstrative and are link to the they possibly not those. Relative clauses The assumed to a proposed Walker (1989) is explicitly realized. segmentation in in this way, heuristic number of by CFs in adopted. our With set the utterances and corpus is the parameters Table 1. as shown Computational Linguistics Volume 30, Number 3 Table 2 Utterances and CBs with the vanilla instantiation. Museum Pharmaceutical Total Percentage [BAR] Number of times at least one CF(U n ) is realized in U n + 1 195 162 357 (35.4%) [BAR] Utterances that satisfy Constraint 1 (have exactly one CB) 189 157 346 (34.4%) [BAR] Utterances with more than one CB 6 5 11 (1.1%) [BAR] Utterances without a CB but are segment boundary 67 96 163 (16.2%) [BAR] Utterances without a CB 168 319 487 (48.4%) 4.1.1 Constraint one/at most 1. one The CB) statistics in relevant Table 2. to Constraint 1 (Weak shown This 1 (that table clearly indicates utterances have exactly of Constraint are C1) is with that the weak version Even without likely to be Weak verified 1,007 counting segment boundaries, C1 is the vanilla instantiation. out of (82.7%) only 11 (1.1%): The verified by 833 and chance that Weak utterances not with violated 11 different indicate sample is by p ≤ 0 . 001 by the sign test. (We will C1 will write hold +833, a to numbers of violators.) On the other henceforth − strong of C1—that verifiers and utterance exactly one CB—is not hand, the with version instantiation: In our every only has 346 (34.4%) likely to hold one CB, whereas this 498 (49.4%) corpus, zero or utterances more one CB. have error have − is the chance rejecting 10%. the null than With exactly 498, of utterances in hypothesis if The of error that Strong C1 +346, obviously much doesn’t chance doesn’t go below 10% hold we count 163 higher than that do not contain references to CFs introduced in even the utterances utterance words, segment if boundaries and therefore are not the previous In other but are instantiation were governed way of by the constraint. we would the vanilla in the “right” in our setting the parameters, have to conclude that CB, the genres contained are very to have a unique but entity coherence does not corpus, utterances major would role in likely that a text is, is coherent: would only 35.4% of utterances in our play a ensuring in finite that contain an explicit mention corpus be “entity-coherent,” to an entity realized The the previous C1 with example instantiation. illustrates clause. following If we why identify there are so many with violations finite of Strong the vanilla in in (9) up into five utterances sentences utterances, and only the last of clauses, the two break of introduced these in (u1). can be considered any sense to directly refer to the set egg vases 20 (9) (u1) These “egg vases” are of exceptional quality: (u2) basketwork bases ... Clearly, there identified are two with ways of “fixing” in which this problem. we would One is to claim only that utterances are in best one sentences, The case other is have indirect two utterances this example, for each sentence. to allow for realization: Poesio et al. Centering: A Parametric Theory (u2)–(u4) CB if all contain indirect implicit is references to Both the egg vases, and therefore all will below. have a The realization 11 allowed. Weak C1 well) is (1.1%) in possibilities are considered fact that utterances worth the corpus noticing. have The more than one CB (i.e., is they violate in as also Joshi, is only reason for the be a order violation “classic” required partial (see, that centering, Prince ranking 1998a 3), when to CFs with e.g., so the same rank in U Walker, and page 21 XML in U + 1 , both become the CB. This two is illustrated in (10), where we i are both realized markup i of show the so that the attributes elements are visible: (10) <unit finite=‘finite-yes’ id=‘u227’>

<ne id=‘ne546’ gf=‘subj’>The drawing of

<ne id=‘ne547’ gf=‘np-compl’>the corner cupboard</ne></ne>

<unit finite=‘no-finite’ id=‘u228’>, or more probably

<ne id=‘ne548’ gf=‘no-gf’> an engraving of

<ne id=‘ne549’ gf=‘np-compl’> it </ne></ne>

</unit>,

must have caught

<ne id=‘ne550’ gf=‘obj’>

<ne id=‘ne551’ gf=‘gen’>Branicki’s </ne> attention</ne>

</unit>

<unit id="u229" finite="finite-yes">

<ne gf="subj" id="ne552">Dubois</ne> was commissioned through

<ne gf="adjunct" id="ne553"> a Warsaw dealer </ne>

<unit id="u230" finite="finite-no"> to construct

<ne gf="obj" id="ne554"> the cabinet </ne>

for<ne gf="adjunct" id="ne555">the Polish aristocrat</ne>

</unit>

</unit>

<ante current=‘ne554’ rel=‘ident’><anchor antecedent=‘ne549’>

</anchor></ante>

<ante current=‘ne555’ rel=‘ident’><anchor antecedent=‘ne551’>

</anchor></ante> In this example, (u229): two discourse entities (realized introduced in (u227) in utterance utterance the cupboard by (ne547) (u227) are and (ne549) realized in 22 (u229) As (ne554)) corner Branicki (realized in (u227) by (ne551), and in proposed (u229) by (ne555)). and in by and their ( grammatical , functions -complement, are equivalent et np-compl for and , under the ranking by Grosz al., available CBs (u229). on the The Web np site), these two CFs gen for genitive—see have the same rank in (u227), the annotation manual of problem occurs with NPs, so of they which are both same This problem with coordinated instantiation both have the same grammatical function. the vanilla be requiring the ranking function total order, which is can also “fixed” by to be other disambiguation factor such as linear order, a as done by Strube easily and Hahn. done by On adding a the in hand, the requirement that one ranking might be total has not been previously discussed only as the one the centering literature, and argue conversely that examples such one CB. We above are arguments issue in againts Section centering’s 5. claim that utterances have return to this Computational Linguistics Volume 30, Number 3 Table 3 Transition statistics for the Brennan et al. version of Rule 2. [BAR] Museum Pharmaceutical Total (Percentages) [BAR] Establishment [BAR] Continuation [BAR] Retain Smooth Shift [BAR] Rough Shift [BAR] ZERO [BAR] NULL [BAR] Total 4.1.2 in Table Rule 3. 2. The The most statistics obvious relevant for Brennan et al.’s version of Rule most suggested by this table is 2 are shown in our consideration ones not that the three frequent in transitions Centering corpus or are that either have without most have been discussed only in been previously discussed the literature CB frequent with transition one, (47.9% without of the total) CB. is (Examples NULL: a limited way. By far the following (u5) (9).) a a also a include up We second (u3), an utterance in only in Passonneau (1998). The (u4), and most found this (18.8%) transition is Kameyama’s discussed Establishment (the second common without transition CB one with CB CB), one followed without), by never its transition between an utterance (between a and with a reverse, mentioned the ZERO transition (An an utterance of ZERO is (u2) a in and (9).) If we ignore NULL, Establishments, in the literature. and ordered ZEROs, example the preferences (Siegal are roughly Castellan as predicted 1988, by Brennan 184–188) et indicates al.: The Page test for of alternatives .001 and pages likely. But only a chance less than CON that RET/SSH, the four transitions are SSH equally more between and RSH, the differences between Shifts and (SSH RSH) and Retains. are significant; and there are + than 23 Grosz et al.’s formulation of Rule 2 in terms of sequences also roughly holds, except that in there Table are 4. too As we’ll few sequences for in Section the results 5, in to our be considered significant, as shown see again which indicate avoiding maintaining is corpus there by CB the numbers seems be repetition; this tendency confirmed or in to a preference for the maintaining table, it in a most dispreference for the same EST/CON position, at least at the of Continuations. level of for finite too long for the salient clauses: sequences Retaining are twice as common as Shifts, sequences figures not As the for do lend much for claim that idea: Retains transitions Shifts, are more prepare the support to the other frequently Of almost other as frequently of by Rule 2, Retains. followed by Continuations than by and the transition over formulations ones proposed the version by Strube based on and Hahn a preference is not for cheap pairs expensive in instantiation—which is not, we verified with the ranking function used the vanilla should em[BAR] Poesio et al. Centering: A Parametric Theory Table Rule 4 2 statistics considering sequences of transitions. Museum Pharmaceutical Total Continuation Sequences Continuation/Retain Establishment/Continuation 17 Retain Sequences Retain/Continuation Retain/Smooth Shift Retain/Rough Shift Smooth Shift Sequences Rough Shift Sequences Null Sequences Other Table Cheap 5 and expensive transitions with the vanilla instantiation. Museum Pharmaceutical Total Cheap transitions Expensive transitions Cheap transition pairs Expensive transition pairs 162 234 396 phasize, the one assumed we find by Strube 396 and of Hahn themselves. 24 Ignoring 35 the 225 of segment boundary utterances, in Table 5. These pairs expensive mean transitions in only and 139 pairs of with that instantiation; cases out cheap transitions, 357 CB(U (the as number shown statistics utterances this see Table 2), of i ) is total by CP(U entity-coherent of predicted ones, i − 1 ). We in do find that of 219 Kibble utterances, (2001)—that the majority are the sense is, (61.3%) entity-coherent CP. “salient” their CB is the same as their in 4.1.3 Table Salience 6. As noted and Pronominalization. our The statistics 217 for of pronominalization pronouns, are shown above, 78 corpus contains In uses instantiation third-person we R1-pronouns 23 demonstratives, and complementizers. 25 to include only pronouns this and complementizers, for total of 295 take If we identify personal with finite 61 pronouns a utterances personal (28.1%) R1-pronouns. clauses, have their Computational Linguistics Volume 30, Number 3 Table 6 R1-pronouns in the corpus with the vanilla instantiation. [BAR] Museum Pharmaceutical Total [BAR] Total number of R1-pronouns [BAR] Number of personal pronouns 144 Number of complementizers Table 7 cb s and pronominalization with the vanilla instantiation. Museum Pharmaceuticla Total (Percentage) [BAR] Total number of realizations of CBs [BAR] Total number of CBs realized as R1-pronouns [BAR] CBs realized as personal pronouns [BAR] CBs realized as complementizers [BAR] CBs not realized as R1-pronouns [BAR] Total number of R1-pronouns that do not realize CBs 58 [BAR] Personal pronouns that do not realize CBs [BAR] Complementizers that do not realize CBs antecedent man Poesio in the 1998) same whose utterance, and 28 is (13%) in neither are “long-distance nor pronouns” (Hitze- and antecedent the same the previous utter- ance. instantiation Table 7 shows is not that the relation between Only pronominalization 55.1% of 374 and mentions CB with of the CBs vanilla straightforward: the 26 pronominalized. ily R1-pronouns And if grounds relative clause complementizers were not included are (on the decision to use a complementizer is primar- among the the that dictated by non-R1 grammatical, pronouns rather (171, than 44.9%) discourse, considerations), R1-pronouns (137, more 35.9%). CBs other as than as On would be realized 73% of R1-pronouns CB. the hand, do refer to the 27 Table 8 analyzes Given pronominalization figures in in Table terms 7, of it the three versions of Rule 1 we are stronger considering. 28 of Rule 1 we the pronominalize should already CB—generalizing be clear that the proposal version Gordon, 3—is Grosz, considered, and Gillion always by In (1993) the the CB other Constraint not fact, 55% to of the less restrictive definition utterances violate it. The of given by of Rule 1 we verified: 1 (GJW 83), pronominalize two versions are considering, however—Rule Poesio et al. Centering: A Parametric Theory Table 8 Evaluation of the different versions of Rule 1 with the vanilla instantiation. [BAR] Museum Pharmaceutical Total (Percentage) [BAR] GJW 95–utterances that satisfy 130 135 265 (96.7%) [BAR] GJW 95—utterances that violate 7 2 9 (3.3%) [BAR] GJW 83—utterances that satisfy 117 105 222 (81%) GJW 83—utterances that violate 20 32 52 (19.0%) [BAR] Gordon—utterances that satisfy 77 45 122 (44.5%) [BAR] Gordon—utterances that violate 60 92 152 (55.5%) the 95), CB if it’s pronominalize the same as CB the if CB of the previous is—are utterance, and of of Rule 1 (GJW verified 95): by most especially Rule 1 (GJW There the anything else utterances. pronouns are two classes violations In (11), CB(u3), possessive pronouns proper noun, to “global topics”. PRODUCT-X , is and referring whereas pronoun is intrasententially realized as a baby. a possessive used to refer to 29 the (11) (u1) PRODUCT-X Infants and children must not be treated continuously with (u2) because it for may long periods reduce the activity of the adrenal glands, and so lower (u3) PRODUCT-X Similar resistance to disease. effects its on mother a baby may occur after (u4) when by is during the last weeks extensive of use of or baby. pregnancy she breastfeeding the In of the pharmaceutical when pronouns leaflets, several violations of Rule 1 are sometimes to realize the product found toward the end text, not For mentioned it in in are used described by the leaflet. example, the following example refers to the cream discussed by the leaflet, the previous two utterances. (12) (u1) A child of of 4 years needs about a third not normally of the adult amount. more (u2) five A course (u3) treatment for a child should it last unless has told you for longer. than days your doctor to use What ize we seem to observe of here is a conflict between pronoun, the “global” preference to real- pronominalize the “main character” most a discourse entity, as a identified and the “local” CB. preference as the 30 product mentioned by number By to the locally salient a the discussed in the leaflet has been of the end of leaflet, so We it is Table enough 7 to justify pronominalization only 9 even when a of Rule it is not 1 from in CF times, that salient Joshi, in that although there are violations Grosz, list. saw and Weinstein (1995), 81 R1-pronouns do not realize CBs. The majority of the Computational Linguistics Volume 30, Number 3 72 cases of pronouns (1) that do R1-pronouns not refer in in to the CB but do utterances which CB without not violate is a CB (the Rule 1 pronominalized majority) fall into two classes: R1-pronouns used in and well—as in (2) used utterances (by in which the as the following example, pronoun both the microscope in and the amateur scientist (u2): are realized a personal and a relative trace) the relative clause (13) (u1) (u2) This who microscope would belonged it to an amateur world. used to explore the mysteries scientist, have of the natural 82.6% one of would demonstrative on pronouns of, in the corpus not Gundel, do Hedberg, realize the CB, which is Passonneau the basis for example, and Zacharski what expect (1993). not This suggests that with treating demonstrative Rule 1. pronouns (1993), and pronouns would improvements On other as R1- only lead 23 to pronouns respect in to the because the corpus, such change would hand, there are demonstrative be affect the results. And indeed, with broader a R1-pronoun unlikely to drastically (GJW 95) includes (11 instead demonstrative of 9) pronouns, we find a of Rule a few 1 more definition of Rule 1 that instead of 152) of Rule 1 (GJW a few less violations (Gordon violations of and 83) (50 instead of 52), none of et al.) (148 is The and “narrow” reported in the rest of but the article are all obtained these differences significant. results of R1-pronoun not include with the definition that does demonstratives. 31 4.1.4 more Differences in Among with the Domains. of The texts in the museum agreement the theory than the texts in domain seem to be tical This is the predictions the for Rule 1. There the pharmaceu- in domain. in especially case (73 of 1,308 CFs, are fewer of texts the pharmaceutical domain [5%], personal as opposed pronouns the 1,731 [8%] museum whereas in museum to (85/211) domain), and the domain 40.3% 144 R1-pronouns), of CB for the in realizations are done via personal only pronouns 29.4% (48/163) (65.4% if we (41.7% consider R1- all pronouns). The the pharmaceutical of domain much museum utterances satisfying (44.0%, 189/430) the strong are version of Constraint for 1 is percentage one (27.2%, higher 157/577), in the domain of than is and the percentage in utterances with in no the CB pharmaceuti- cal much (55.3%, 319/577) that are not segment first (39.1%, boundaries 168/430). Finally, higher this 72% second of domain or almost utterances in than in the main NULL ZERO (415/577), whereas 54.6% the pharmaceutical in museum do- are (235/430); transitions of EST CON is just are 126/577 and These also higher in the domain (133/430, 31%, the percentage [21.8%]). in the museum domain number of versus pronouns in differences are part due to the large Constraint second-person 1 improve if we you the pharmaceutical domain, so that pronouns the statistics CFs, for we will below. A treat the entities referred to by much important see role in second reason for these differences is these as as more providing that layout plays way a of (See Section the pharmaceutical 5.) domain, a different achieving coherence. 4.2 We Varying now the Utterance Parameters begin to explore most full of discussion of alternative instantiations. parameter settings. prevent with all the In As always, this article, we space constraints a only discuss variants briefly and analyze length the instantiation the results the quite at some Poesio et al. Centering: A Parametric Theory that The identifies utterances with sentences; the on results are summarized extended report available the companion Web with graphs. technical of of interested site contains a more extensive discussion some instantiations on the variants; Web In readers to the this subsection we are also encouraged try further of (parameter site. ) of consider how the definition utterance uttdef and the value the parameter previous utterance affect the claims. 4.2.1 on Treating Coordinated or Verb NLG Phrases as Utterances. of Many researchers is, that each element a coordinated VP working spoken dialogues assume as a separate , utterance: VP in counts that that We should send is the engine to Avon and hook it to the Treating tanker car the coordinate more VPs hook it to the tanker 1,007), separate which of in our car actually would corpus would a separate as utterances of utterance. in coordinated (1,041 were course lead to worse course result utterances vs. implicit If we results we unless these utterances (but treated as containing nonsignificant better results Strong an C1 (48% trace. do violations instead so, slightly for of obtain significantly) and Continuations Rule 1 number and of Rule 2 (with 49%), differences for Retains). slightly higher numbers of and slightly lower 4.2.2 of Using All Verbed of Clauses is Instead of Just the Finite Ones. with A second to-clause in as utterances all clauses a verb, including, extension the definition instantiation, infinitival utterance to treat for example, the well, on John our wants to visit Bill . The results With with such an we as of crucially many depend more grammatical (1,267 instead assumptions. a get course utterances of 1,007), most such which, definition, of the example in infinitival clause just given, if we do not contain explicit mentions of like the argument in subject we would position; find so again, more didn’t assume of Strong that traces present clauses, significantly violations C1 (685 are instead of 498). such of Using matrix a crude mechanism nonfinite for tracking traces (adding a trace to subject number the clause to all complement clauses), we referring the of (598) with instantiation, still find a larger number of violations is much greater, than the vanilla but because the utterances of (47% instead of these of 1. As Rule 49%). violations 2, We find no represent a significantly number the total significant in differences in lower percentage NULL of Rule the violations (45.0% instead of for 47.9%) this change results 18.8%) SSH (5.6% instead of and significantly more EST significantly (22.1% instead fewer of transitions and 3.7%). 4.2.3 Restricting Finite Clauses. of In general, the best results for C1 are number obtained of by considering In larger particular, chunks text as a single utterance, finite occur are obtained thus utterances. violations by (as not reducing the fewer ances clauses parentheticals, subjects in considering as utter- that as ), matrix as to was surprise me with That John could [BAR] big and clauses subject (as do in this Mary a to as an empty Computational Linguistics Volume 30, Number 3 4.2.4 Relative problems we Clauses. Relative The clauses may turned out to be one (without face. reader recall that Kameyama of the most complex had to tentatively proposes empirical support) that relative clauses have a “mixed” status: They should be locally merged treated with as updating the local focus, This but at the proposal, global however, level they should be the embedding may utterance. with of seems to assume that the local focus be first updated processed, the which content is certain utterances some time after they have of been In study, we a rather radical change to the basic assumptions in which the framework. this simply with compared one in which an instantiation we relative clauses are treated as utterances (i.e., not they are not. In addition, considered treating (embedded). relative clauses as adjuncts The figures as embedded) were obtained and treating them as complements 32 reported so by clauses and akin adjuncts. 33 Not far treating relative as utterances in 6.5% as to in number of treating relative clauses with as separate utterances (908 results instead of a 972) reduction in the to fewer violations of Strong utterances respect vanilla − [439 without CB, 13 with CBs] and instead of 469 [457 12]); C1 (452 utterances Rule of a violations is two higher (49.7% vs. 48.2%). The number and of however, the percentage 1 (8 [2.7%]). From of of Rule 2, violations EST, number point view a lot of of also stays the same the seem their goes down by almost 15%, relative clauses 17.3% (from 190 to function 157); we as since 30% in SSH increase in NULL, to 50.6% of to Everything also see a reduction and an to In the numerical total. else not stays the same. would purely not improve terms, then, Furthermore, treating relative most clauses important, as separate finite and we utterances the results. would make it feel difficult that not treating maintain relative clauses as separate utterances in principle of that utterances we will are identified with finite very clauses. For to the these reasons, the rest the article continue to count relative clauses as finite clauses. Suri 4.2.5 Suri McCoy and McCoy’s (1994) Definition of Previous after- Utterance. before-clauses As discussed more in Section 2, and (i.e., suggested that and ones, behave Cooreman like em- bedding Sanford elements (1996) like complements) than like coordinating and and The found evidence supporting of our this treatment for when clauses, with previous utterance When parameter is script can be used to compare this proposal as well. Kameyama’s. not this parameter in (14), set “Kameyama-style,” adjunct (u3) is clauses as that, the previous utterance for (u2). (This are treated was embedded, so (generalized) the setting Suri used for McCoy, the results discussed so far.) When the parameter is set to and (u3) is (u1): adjunct clauses are treated as embedded, so that the previous utterance for (14) (u1) (u3) John He woke up (u2) when Bill rang the door. had forgotten the appointment. Using nificant Suri and McCoy’s in number definition of of previous of utterance Strong C1, results in in a improvements small but sig- reduction concerning Strong R1 C1 (GJW is 95), the in violations and 20 small, not significant, Strong improvements small but C1 with for Kameyama’s Rule 2. As far as concerned, utterances that violate John ... Poesio et al. Centering: A Parametric Theory definition satisfy it +20, under 9, Suri and 03). McCoy’s, The but 9 is utterances not, become C1 test, − p ≤ . With we however, violations the reduction 1, find sufficient (by sign (+355, 458). Rule number of for Strong to be verified − GJW 95 increases (+287), that the number utterances that verify the version (8 [about 2%]). but the and percentage of violations We stays the should of note, same however, that the differences only have Kameyama mostly between not to do with Kameyama’s Suri a type of and Suri’s definition previous utterance McCoy: clause that was discussed in (15): briefly by and at all by and relative clauses, as (15) (u1) (u2) This brooch which is is one made of of titanium, (u3) It was made Anne-Marie the refractory metals. by Shillitoe, an Edinburgh jeweller, in 1991. If the “generalized Kameyama” (u3) is definition of previous (u2); utterance is adopted, utterance the relative clause, this violation of the Strong pre- vious C1. In for Suri McCoy’ instantiation, causes a is the ‘generalized and approach. by contrast, If the relative clause not treated as embedded; this seems or to were be the better would as separate utterances treated as embedded in relative clauses were we treated find equal number of although about 20 both instantiations, in an instantiation. One violations, in which violations would be different with way each is example (7). PRODUCT-Z is the not difference mentioned does in have to do if-clause, the adjuncts are In Suri Strong handled McCoy’s C1 is violated proposal if works (u2) is the adjunct (u3). and therefore taken as previous Kameyama’s. utterance for this case, and better than (7) (u1) (u2) You if should not use PRODUCT-Z (u3) Whilst you are pregnant or breast-feeding. you are receiving PRODUCT-Z .... Conversely, inside, introduces in the following example, patch, the which adjunct is clause, as you may (u2) entity the then referred to in (u3), damage the patch the of C1. In Kameyama’s so treating the adjunct of as embedded leads to a violation this case, definition previous utterance gives the right result: (16) (u1) (u2) Do not may use scissors (u3) as Take you out damage the patch inside. the patch. Given will that Suri these improvements McCoy’s are significant, when if small, is in the rest of the use definition uttdef set to finite clause . However, article, we and idea and may the contrast between (7) and (16), our discussion, especially identified with gives next. utterances be best sentences. We further support to the that consider this setting 4.2.6 is Sentences. which identifies The setting of uttdef with with the most The dramatic impact on Strong C1 that illustrated with (9): utterances If sentences. identified with reasons for this have in in are sentences, there are only already been utterances utterances example, with both containing instantiation references more to the egg vases. The two that is Strong reduction violations such that this utterances verify C1 Computational Linguistics Volume 30, Number 3 Table 9 Statistics relevant to Constraint 1 when utterances are identified with sentences. Museum Pharmaceutical Total (Percentage) [BAR] Number of times at least one CF(Un) is realized in Un+1 131 147 278 (41.6%) [BAR] Utterances that satisfy Constraint 1 (have exactly one CB) 126 138 264 (39.5%) [BAR] Utterances with more than one CB 5 9 14 (2.1%) [BAR] Utterances without a CB but segment boundary 65 80 145 (21.7%) [BAR] Utterances without a CB 75 171 246 (36.8%) than violate it, although not so many as to ensure verification at the 5% level. 34 statistics Constraint 1 with this definition of shown in Table 9. The Although relevant Strong to C1 is not if we utterance number of verified consider Strong all 669 are NPs, still C1 (264) segments is of text that contain number the of utterances (260). that satisfy slightly larger than However, the identifying those that don’t The main utterances 1 these is with sentences that the number also of has several violations of negative Rule 1 (if small) effects. In of Rule among (GJW 95), 50%, from 8 to 12. The increase goes up: in the case more by CB; in reason for this CB’s utterances have a but some cases, the problem is part simply that be viewed not being updated quickly enough. Consider could as the the following example: The not pronoun if it in sentence (s2) violates This Rule is 1 if utterances in first are viewed are as clauses. because the case (s2) as sentences, but CB, they , viewed whereas with instantiation (cl3) has is a single the wall lights the vanilla − utterance, with Rule CB (GJW one of these 95) wall Rule lights 1 . Because (GJW 83) the number clause, of violations is a separate +209, respectively, 1 +287, are still +243, verified (+252, still − 55, 52, with 12; quite small, both − opposed and as 8 and to − and − the vanilla − instantiation, Suri setting), 35 although Rule 1 (Gordon et al.) still isn’t (+97, − Note (129) also much that with this instantiation, number the number of CBs realized by R1-pronouns 167). is smaller than the realized by other types of NPs (209). Poesio et al. Centering: A Parametric Theory Table Rule 10 2 statistics with sentences as utterances. Museum Pharmaceutical Total (Percentage) [BAR] Establishments [BAR] Continuations [BAR] Retain [BAR] Smooth Shift [BAR] Rough Shift [BAR] ZERO [BAR] NULL The Table we results observe for Rule 2 Note more are not much Continuations that different from those obtained number and fewer of Rough NULLs. Shifts The with finite clauses, but in do 10. statistics of Smooth are shown the greater Brennan is than Page Shifts, although There the ranking suggested by et al. still verified by Grosz still too few sequences to truly test the version of the Rule 2 proposed test. are by 2 proposed et al., Strube but the preferences Hahn, are roughly 10 verified. many As for the version of Rule (191) by and ones there (18). still times as expensive-expensive sequences as cheap-cheap 4.2.7 Interim Summary. on Strong The C1 effect of Rule the 1 changes in the definition in Figure of 1 utterance utterance summarized and Figure and previous respectively. As figures show, and most are 2, the often such The one changes have fairly is identifying small effects, even though with the effects are significant. exception we all clauses as utterances implicit also has a positive of impact, utterances sentences; treating assume nonfinite provided that matrix clauses contain an realization the subject of the Strong Even clause. C1, though we will identifying not utterances with sentences leads to much better with abandon the hypothesis that utterances may results for finite of This simply other is in in coincide clauses. which part for theoretical reasons, such as RST, the fact that in theories finite discourse Also, identifying “units” are assumed, with such as these units are generally increases clauses. in number of utterances of Rule sentences 1 (from 8 leads in to small, but instantiation significant, will [2.8%] 12 [4.5%]) the moment and in violations the number other of Rough Shifts (from the 2.9% vanilla 4.6%). We to see in that there are ways of to instantiation also Strong a necessary. satisfy C1, so identifying utterances with changing the vanilla that sentences is not strictly In the rest of on the instantiations article we will, in which therefore, study the identified effect of on in utterances which are identified with changes = instantiations with finite to other parameters (henceforth, both (u s). f) clauses u and they are sentences = 4.3 Realization In this section we discuss the effect . of changes in the values of the realization param- eters: realization and CF-filter Computational Linguistics Volume 30, Number 3 Figure The 1 effect of utterance parameters on Strong C1: A summary. 4.3.1 way IF: Indirect Realization number of + u = f of . Strong Examples Constraint such as 1 (9) is to indicate that reduce violations allow for indirect another to realization: make Then the CB the of bridging references And to the indeed, egg vases if we modify in (u2), (u3), and (u4) would them instantiations the (vanilla these utterances. , our of Suri the “best” McCoy’s among proposal the u = f − using generalization to Strong the previous C1 is with to allow 525 for indirect and determine utterance) (54.0%) realization, the reduction 325 zero such that utterances having exactly one CB in violations to and (+525, having 325). or more than one (33.4%), Strong C1 is verified by the sign test 36 − Poesio et al. Centering: A Parametric Theory Table Rule 11 2 statistics with indirect realization, 4 u = f. Museum Pharmaceutical Total (Percentage) [BAR] Establishments [BAR] Continuations [BAR] Retain [BAR] Smooth Shift [BAR] Rough Shift [BAR] ZERO [BAR] NULL However, allowing for indirect The realization first negative has a negative is effect on other just with change to u=s does. effect that the number of claims, as the ances more one CB doubles, This from 13 with utter- McCoy” instantiation than (1.3%) almost 22 (2.3%). is the increasing “generalized Suri of because by the number and in we The increase to “persistent entities,” number the chance of of their of having Rule 1 an equivalent ranking 8 with the previous Suri utterance. McCoy instantiation violations 16. But with exactly indirect doubles: from more the and Rule 1 utterances increases, have a from CB, 295 the number 467, of to because utterances that matter realization not to so that the percentage of for the purposes violations to Rule of also (GJW change opposed that much. 2.7% With indirect with realization, Suri 3.4% of utterances violate Rule 1 does 1 Rule 95), 1 as (GJW 95) (+451, to 16) generalized Rule 1 (GJW 83) and (+318, direct 149) realization. As a result, − Rule (Gordon isn’t and − whereas 1 (+136, 331). An of are pronoun still verified, et al.) still − in (18). violation The of Rule 1 (GJW in 95) if we CFs example NP (u42) allow to be indirectly a that becomes a realized is shown introduced one CB, NP stand realizes is not pronominalized. in (u39), a bridging entity the two stands which is reference to the discourse by its the it therefore realized in (u42), and thus becomes but (18) (u39) originally The two stands are of the same date as the coffers, but were (u42) designed make it was to hold rectangular in cabinets. One stand adapted other. the late 1700s or early 1800s century to [BAR] the same height as the Finally, in the Table change 11. to On indirect realization has number a big impact of NULL on the statistics for Rule 2, shown (to the 30%), positive side, the of transitions goes down significantly However, less than increases and the percentages in number the of RET four RSH 2.6% the greatest are the (from “classic” 3.8% transitions go up. more (from to SSH 10.0%). The fact that there are many more RET than CON to 13.1%) and many (BFP) RSH means that this is the first instantiation for which Rule and implicit is than verified by implicit a Page NP test. modifiers The reason (i.e., for this can be seen in (18): Because 2 not realizations ), are never CPs of (Rule one 2 stand [Strube is interpreted Hahn] as one isn’t of the two stands they are of an utterance. increases from and 207/747 percentage cheap transitions 154/747, still verified, although [27.7%]). the [20.6%] to Computational Linguistics Volume 30, Number 3 Table 12 Statistics about Strong C1 with u = s and indirect realization MUSEUM PHARMA TOTAL [BAR] Number of times at least one CF(U n ) is realized in U n + 1 [BAR] Utterances that satisfy Constraint 1 (have exactly one CB) [BAR] Utterances with more than one CB [BAR] Utterances without a CB that are segment boundaries [BAR] Utterances without CB Below, we indicate the instantiations DF with u=f, Suri with and McCoy-style treatment of adjuncts, and direct IF. realization as and those the same settings, but indirect realization, as 4.3.2 1 IS: Indirect Realization 12, if indirect + u = s . As instantiation (henceforth, is one might Table better realization combined IS) 390 with expect, the results for Constraint in get even With the u=s out setting. of 669 As (58.3%) shown Strong C1, this 177 (26.5%) it—significantly utterances satisfy with and (henceforth, violate DS). On better more CB the (and other than the instantiation u=s and with direct realization one with hand, the number instantiation), doubles again respect to the DS of utterances number of 26 (3.9%) than from violations of Rule 14 (2.1%). almost R1-pronoun). DS The to instantiation, from number 12 (4.5%) 1 of to 26 (GJW (6.7% 95), of as well, the 390 doubles again isn’t utterances invalidate with with Rule a CB respect to the While 1 and (GJW an 95) (+364, 26), we it is this Rule 1 (Gordon number violations of enough − times the violations with with to As what three the instantiation vanilla instantiation. 75% for of called it: +97, 293. et al.), even this more than utterances violate 37 − Just The in results with Rule Rule 2 2 (BFP) are comparable is not to those obtained with as verified according to Page the IF instantiation. is that case, in number of NULL (to a of [9.4%]) is a great reduction the transitions 23.2%). test, The even though there RET even is of Rough than with Shifts IF (114 [14.9%], [17.0%], percentage greater (100 almost twice the percentage of CON Smooth as Shifts that [5.2%]). If we ignore almost three times the percentage 30.6% boundaries, cheap transitions are 136/444, of of the total (as opposed to 22.0% segment with DS and 27.7% with IF). 4.3.3 Second-Person CFs. Second-person pronouns (henceforth, assumed than anaphorically (see, PRO2s) e.g., Di Eugenio are generally to be used deictically rather 1998). Poesio et al. Centering: A Parametric Theory However, it has been CFs suggested (Byron in recent Stent work 1998). that In especially our in dialogue, they realize corpus, especially in may actually and 38 maceutical in PRO2s are very numerous of and often and the phar- role domain, maintaining And seem in to play an important the of coherence CFs the discourse. count reduce the number of fact, allowing violations of PRO2s Strong C1 to as realizations does instantiations of theory, with with both the u = indirect f and the Even u = with s DF (and the Suri-McCoy both of direct and with realization. museum PRO2s the to not count as CFs is sufficient setting the in on its previous own utter- ance parameter), Strong C1: The allowing number of is to verify domain Strong affected, C1 increases but the from pharmaceutical 164 273, domain, the 464 utterances that C1 satisfy With 367 DS, if violate we it, which makes to so that in total (by utterances satisfy 03). and PRO2s CFs, the 332 constraint verified Strong the C1, sign test, CFs). 214 p ≤ . don’t (as opposed to +264, treat − 242, indirect 260 when as Allowing PRO2s utterances are not verify and realization we get even better results: With IF, we treated get +623, as for − PRO2s; a with significant IS, improvement we +439, even 145. over the instantiation with direct realization and get − The of NULL results with Rule 2 is are also improved (for by age reduced DF, treating PRO2s down to 35.0% as CFs. The [from 47.6%]; percent- DS, As 30.8% [from transitions 40.8%]; greatly for of IF, to 18.2% [from 29.5%]; (Kibble for IS, 2001)—EST, to 15.1% [from CON, 23.2%]). for to SSH, a result, In RSH—increases. the percentage However, “continuous” RSH transitions and SSH increase, as well as EST and CON: RET, and in the IS IF instantiation, instantiation with RSH PRO2s, (18.4%). EST Because are the most of common increases, transition (20.2%), of not are these treating PRO2s but the isn’t CFs IF fix the problem with rule 2 (BFP) observed as realizations Rule does (Strube with and IS. There are no significant changes with above: Rule the still Hahn). verified 2 The and results pronouns with Rule 1 R1-pronouns. crucially depend In on whether we do PRO2s or do not consider second-person in more as of Rule 1 (GJW 95), either in case, letting in realize CFs results more violations if we have a CB both absolute and PRO2s and therefore R1-pronouns, count as violations increase or relative terms, because But utterances verifications the don’t then in of is rule. (3.4%) DF, from 8 consider (3.5%); (2.7%) 12 IS, (2.9%); as from 26 for (6.7%) DS, from 31 12 (7.1%). (4.5%) the from small: 16 20 If to we 17 (5.1%); violations for to PRO2s for IF, R1-pronouns, however, to and for of of to instantiations violations ones: Rule 31 1 (GJW 95) treat as the percentage almost DS doubles for the u=s violations for DF (7.6%), triples for the 1 (11.4%), u=f 95] 51 for IF (9.1%), and and in 67 for is (15.3%). (Of The course, in all of 38 Rule for [GJW these remains sense.) reason for this increase cases, in violations is that PRO2s verified do not a statistical of PRO2s not seem of CBs, to be in very instantiations. good indicators Given of the CB: About half it are realizations whether that it’s not a idea all PRO2s as R1-pronouns; these seems good to treat it’s results, clear to us clear realizing As we find the position that PRO2s less to treat them as in of cf s. indicate rest where the article, we would will not include play a deictic function among CFs, convincing, we will the doing so result in major their referents The interested the but is on Web differences. reader advised to try the alternatives the companion site. 4.3.4 Predicative NPs. The in two increase alternative in views number considered of CFs. so What far about if we which tities result the were en- to realize both an to [BAR] Computational Linguistics Volume 30, Number 3 attempt of to reduce modification the number of CFs negative instead? impact Prima on C1, facie, one would imagine Rule might to have a disappear. but perhaps some of this type of 1 the violations Among the NPs of are NPs, that might in is, be NPs thought not to introduce that of policeman in CFs, an obvious candidate predicative like a John But in is a policeman that our play the role predicates were instructed the logical mark form an utterance. fact, because annotators to in up John rather than filtering a policeman as antecedent NPs of on contrary, away such does not subsequent anaphoric expression these examples, impact all; the it does have a significant (if have small) negative any positive on result NP Strong at in C1, 39 because in some cases the annotators of had been forced to mark up an instructions. predicative Two position as the antecedent is such examples are given below. an Especially anaphoric marked in expression against the not the second of Bjorg: case, it clear how else the annotators could have the antecedent (19) In the following we will continue to treat predicative NPs as not introducing CFs. 4.3.5 Interim Summary. The realization on Strong parameters C1 have Rule an even 2. Either greater impact than the indirect utterance parameters, or especially pronouns and introducing CFs, allowing Strong treating When second-person as is sufficient for realization for C1 to be verified. On the two other settings are combined, in constraint. in the number hand, allowing for indirect a large majority of utterances verifies the overall of increases the of violations of Rule 1, realization also results significant In although the percentage such increase violations in remains number pretty of RET small. CON an the and RSH addition, indirect realization leads to instantiations RET such and with fewer SSH than RSH, and Rule 2 (BFP) is not that there are fewer of than indirect we Treating verified PRO2s by any the as on Rule of CFs, 2, while when sufficient realization we Strong have to cause C1 discussed. realizations with to be IS verified, instantiations, has less an but combine this setting the we of obtain effect an instantiation in which RSH is in the Figures most 3 common 4. transition. The effects of the realization choices are summarized and 4.4 Ranking 4.4.1 Grammatical Function + Linear not Disambiguation. We most observed above that cause when function does always specify a unique, highly ranked CF, be- grammatical which using that ranking function of weak some utterances of Constraint end up with 1 more than causes violations the version seen above (up one to 5.7% CB, of the total however, in the the IF instantiation problem treating in is can be fixed PRO2s as CF realizations). We also men- tioned, order, which, that this by requiring the ranking factor. function a easily done by adding tie-breaking Given to be total in Gernsbacher turn, Hargreaves (1988) Gordon, a Grosz, Gillion (1993), the results and and and Poesio et al. Centering: A Parametric Theory Figure The 3 effect of the realization parameters on the violations of Strong C1. Figure The 4 effect of the realization parameters on the percentage of violations of Rule 1. the most obvious disambiguating say, factor CF is linear order: And Whenever indeed, two we CFs order the leftmost a higher rank. saw in are Section equally ranked, assign to, is Strube Hahn (1999) in 2 that with linear used by other and to resolve a function than grammatical function. In ties, albeit this section we conjunction ranking we GFTHERELIN. function obtained The by adding linear order evaluate the ranking with to grammatical function, which Table call 40 13. results this ranking function are summarized in Computational Linguistics Volume 30, Number 3 Table 13 Summary of results for .01 indicates Strong C1, Rule a claim is not 1 (GJW 95), and Rule ♠ verified at the .05 2 (BFP) with GFTHERELIN ranking. that level; ♦ that it’s not verified at the level. [BAR] Rule 2 (BFP) (Page test, probability of not being verified) [BAR] Instantiation Strong C1 (Percentage Rule 1 (GJW (Percentage 95) violations) violations) DF-predicate +352, − 450 (46.3%) ♠ +291, − DF-predicate+pro2 +465, 355 (36.5%) +403, 11 15 (3.6%) − − (3.6%) [BAR] .001 .001 DS-predicate +273, − 249 (37.2%) ♠ +259, − DS-predicate+pro2 +347, 197 (29.4%) +325, 14 22 (5.1%) − − (6.3%) [BAR] .001 .001 IF-predicate IF-predicate+pro2 +529, − 310 219 (31.9%) (22.5%) +463, − +635, +325, 18 22 (3.7%) − − (3.7%) [BAR] .05 1 ♠ ♦ IS-predicate+pro2 IS-predicate +408, − 113 157 (16.9%) (23.5%) +432, +378, − +469, 37 30 (7.4%) (7.4%) − − .05 1 ♠ ♦ (including The table PRO2s summarizes without eight instantiations: For DF, instantiation, DS, IF, and IS, each in two variants of Strong and C1 them). Rule 1 (GJW each 95) the table of lists verifiers and well violations of Page and Rule 2. and the percentage violations, as Strong Adopting as the results C1 is GFTHERELIN the test for This as a is ranking function only doesn’t lead from to major changes as because the change the results obtained as far with concerned. CBs simple grammatical function is most from reclassified as in having Strong one, that the C1 and with utterances the instantiations previously classified on DF—the that would as having two get a violations—those based number benefit the multi-CB reduction improvement is 1–2%, instantiations. although this The is enough improvements to make of sentences fairly small, typically with sign test with the significant CF instantiations, by the in since with all sentences it’s more common for more are greater than instantiations one the u=s to in be of which realized PRO2s the same grammatical position; are considered as realizations of CFs, for example, we find in the that 5.7% IS with utterances instantiations, (38/669) have more with than one CB. However, Strong C1 is already verified As these in even simple grammatical worse cases, better results with Strong function. all previous not Rule The 1—although, with DF not instantiations much worse C1 are counterbalanced results again, that Rule 1 by for so ends worse: up For being verified. we find +291, results 11 (3.6%) the with instantiation aren’t not including significantly NPs example, − pronouns, opposed the +280, 9 predicative and second-person as to − with The number of for the same simple violations of instantiation grammatical with function DS instantiation ranking. if PRO2s CF Rule 1 is significantly +325, -22 (6.3%) greater +310, the -17 (5.2%). In of are treated vs. the additional five as realizations: 1, however, whose problem CB is is two violations simply that by adding a disambiguation element, we of Rule the (because more one CF is turn utterances with CB. One undefined is (20): than equally ranked) into utterances a such example (20) (s7) Intended of to hold jewels or with small precious items, the interiors or pewter, of this pair coffers are lined tortoiseshell and brass with Poesio et al. Centering: A Parametric Theory Table 14 Transition percentages for IS with GFTHERELIN ranking. [BAR] Museum Pharmaceutical Total (Percentage) [BAR] Establishments [BAR] Continuations [BAR] Retain Smooth Shift [BAR] Rough Shift [BAR] ZERO [BAR] NULL is With simple grammatical our function, not both the coffers well-defined and brass CB. are of script as having a As CBs of (s8), which pronominalization therefore treated by non-CB, , is not a result, the (s8), however, a premiere with GFTHERELIN, partie marquetry counted as a violation. only CB. becomes a violation since the coffers becomes its icantly With the IF instantiation, the percentage with of violations of Rule 1 (3.7%) greater the simple grammatical function (3.5%). is nonsignif- of than with percentage IS instantiations (7.4%) is worse The percentage .01 violations with the two (6.7% significantly level) grammatical function 7.1%, (at the Table 13 than simple of instantiations of with Rule 2 (BFP). GFTHERELIN and respectively). shows that using Whereas with has a positive effect on the indirect simple 2 grammatical function, Page none number violations Rule (BFP) of GFTHERELIN by the rank test, with the PRO2s IS instantiation realization verifies the CF does (All (although only at the instantiations .05 level), as does IF if are The treated main as realizations. is direct realization the for this change significant reduction in still verify of RSH rule.) with GFTHERLIN, reason a RSH 9.7% especially for with the IF IS, and from IS 14.6% instantiations: 11.4%. With the percentage in from 7.6%; With IF we see a instantiation reduction of to RET RSH is to instantiation, we observed and with about twice what we find with the DS DF the percentage the otherwise just as simple with grammatical DF. With function IF ranking, but instantiations, the we results are pretty similar to increases those in CON RET the and the IS in RSH. In Table get 14, small we but significant and instantiation, and a reduction with other report the complete percentages for this for comparison The GFTHERELIN ranking functions. change to hardly affects the Rule relative 2 (Strube percentages Hahn) of cheap and expensive transitions, so the results concerning and do not change. The IS instantiation without need with GFTHERELIN PRO2s ranking CF is the one in which all three verified though Rule claims are to treat as realizations, even 2 is Computational Linguistics Volume 30, Number 3 verified with this instantiation instantiation when making only at the .05 level. with We will other therefore concentrate on this comparisons the ranking variants. 4.4.2 Linear Order. Among is the one ranking functions CFs in alternative order to simplest that the of grammatical function, perhaps the from the This ranks was their occurrence left ranking function explicitly proposed in the utterance, (1993) to right. in German, of by for by, scrambling others, Gernsbacher and effects order Rambow to account observed facts about Hargreaves of (1988), mention have been Grosz, Gillion (1993), among Stevenson, Crawley, and Kleinman (1994). Gordon, Using and one would order instead and of GFTHERELIN and linear matters has no effect at all on Constraint since the constraint to be verified is 1, as expect, mentioned all that However, in for successive utterances were whether whether discourse is entities are no observed and with Rule the ranking function 95), total. with IS, we find significant This 378, 30, differences with order is order, opposed 1 (GJW either: + − linear as to +377, − with GFTHERELIN. in is linear a very good approximation 31, 41 English: because Subjects which to occur in first position, objects of grammatical function (BFP): The one tend second for RET the differences SSH are significant is in with position, IS, etc. CON claim no at .05 and enough become RSH Rule that Rule 2 is enough become level. (The rule is still verified in the DF 2 DS instantiations.) longer verified even the and the ing All in all, these GFTHERELIN; results are not grounds however, to argue that linear order is a so better rank- function than 42 order (which is because the differences proximation far easier to compute) might are small, they also suggest of that linear be working a good English. ranking for practical applications with ap- grammatical-function 4.4.3 Information Strube Structure. Hahn (1999) Replacing (henceforth, GFTHERELIN STRUBE-HAHN): with the ranking hearer-old function pro- posed more by and inferrables, hearer-old rank en- tities and linear-order not highly than and these different results for Strong higher lead Less C1, than entities—cannot did to was for the reasons already of order—we expected the fact that—again, just as in discussed for ranking. find with Rule 1 (GJW the ther, any significant differences 95), case linear with didn’t IF IS instantiations, we find one more with ei- GFTHERELIN. although This the and mean violation than 43 rather, doesn’t that the exact out”. same We utterances are one violations in in both which cases, STRUBE-HAHN that the differences “balance in of Rule already 1, whereas saw GFTHERELIN example This is ranking first results a ranking which the may in violation doesn’t: sentence in CBs. (10), We illustrating the kind of situations in (21) in include a partial ranking result two repeat that sentence and the preceding sentence: (21) (s67) An inventory of Count Branicki’s possessions objects made at his on death describes both the corner of mounted cupboard Chinese and the displayed its shelves: a collection with porcelain and clocks, some embellished porcelain flowers. [BAR] Poesio et al. Centering: A Parametric Theory Table 15 Transition percentages for IS with STRUBE-HAHN ranking. [BAR] Museum Pharmaceutical Total (Percentage) [BAR] Establishments [BAR] Continuations [BAR] Retain Smooth Shift [BAR] Rough Shift [BAR] ZERO [BAR] NULL (s68) The drawing of it , of must the corner cupboard, or Branicki’s more probably an engraving have caught attention. Branicki, As the corner which cupboard is in NP-modifier is in object position, it which—while gets higher ranking not in (s67) than Count position, explicitly discussed CB (s68), presumably its pronominalization fall among the “other” is cases. As in the centering literature—will is of a Rule result, the With cupboard STRUBE-HAHN the and CB (s68); ranking, Count Branicki Conversely, is predicted by the highest-ranked (22) is entity of in (s67), 1. GFTHERELIN of hence the violation. an example which and therefore the and ranking results in a violation of Rule 1, while STRUBE-HAHN ranking doesn’t: (22) (s88) Paris,’ Christened Andr by his Boulle’s contemporaries name is as ’the most skillful with artisan in è-Charles with marquetry of synonymous pewter, the practice of veneering (s89) Although furniture not invent tortoiseshell, Boulle was its and brass. he did name its the technique, name: Boulle work. greatest practitioner and lent his to common In this example, Andrè-Charles , is Boulle’s name CB , the of subject (s89), of in (s88), which, is however, ranked higher it is than Andrè-Charles pronominalized, Boulle and therefore Boulle the even and the technique he invented are. Notice not (21) and (22) though both that Sidner (1979) are almost stereotypical needed, instances one of the class of examples that two for animated and one led to argue that we foci are issue in Section 5. entities, for the entities The one acted upon; 20% where RET STRUBE-HAHN return to this (BFP). About claim of CON ranking 20% makes of RSH a clear difference SSH. Although is Rule 2 we find Rule more (BFP) RET become CON more and 2 .01 RSH about still and than in SSH, become than cause be verified at the level all instantiations these changes are sufficient to to considered. 44 Even percentages IS, with IS and instantiation STRUBE-HAHN ranking are one presented proposed in Table Hahn—we however—the closest to the by Strube 15. The transition with and still find many more expensive transitions (272) than cheap ones (172) Computational Linguistics Volume 30, Number 3 and ones almost (56), three times Rule 2 as many (Strube expensive-expensive Hahn) is not sequences (137) as cheap-cheap so and verified. 4.4.4 Summary. most Because Strong C1 of is the most problematic claim, would it was to be expected that the studied parameter centering, ranking, It is have nevertheless a smaller utterance realization parameters. interesting impact than the Rule 1 and (GJW More 95) virtually identical under the three versions of that the results we for are Rule 2 (BFP), which is not ranking considered. instantiation differences with with can be linear-order found for any Adopting STRUBE-HAHN ranking and is verified only verified by GFTHERELIN. by a few in instantia- tions of utterances classified into one of ranking does result a greater percentage probability of Rule being 2’s (BFP) Finally, the “continuous” not classes and ina lower were sufficient being falsified. Rule 1 (Gordon even these last to either al.) or changes (Strube parameter Hahn) settings to cause et Rule 2 and to be verified.

In this obtained; section we we first discuss the effects of of different theory, parameter settings on the re- sults make then analyze the claims centering work (empirical draw a few theoretical conclusions, and some suggestions for further and theoretical). 5.1 Setting the Parameters 5.1.1 ways Comparing of instantiating Instantiations. A central goal of this study was to of compare its different centering’s parameters, not and different versions (e.g., claims, whether on a single Brennan data set, also examining Strube of Rule combinations 2 would previously Hahn when considered et al.’s version be verified Our first the interesting parameters were in set as suggested sense is if by and and only Rule parameters 1 (GJW 95) are set Rule in vice versa. the 1 (GJW most 83) “mainstream” way (the result this instantiation) that the The “vanilla” Constraint 1 and negative. As with are clearly instantiation verified. only 35% results concerning are especially is, CF(U ) CF(U this ) (Kibble 2000; Karamanis of utterances are continuous—that n n 1 = 2001)—only the weak one version in of our Constraint 1 is − verified. Strong idea of C1, the best-known formulation, and the Another that only interesting view best observation captures the clearly hold. with more is that one if “entity coherence,” doesn’t CB: ranking The is required of to be partial, is some instantiations. only utterances 1% with end up is the vanilla obvious, instantiation than percentage with such utterances This our but can be as high it as not 6% some perhaps but to knowledge, had been previously discussed. is As for Rule 2, Page with the vanilla instantiation arguably, the most version proposed by Brennan et al. with verified instantiation by a is rank test, but of NULL the striking (47.9%), fact about (18.8%), ZEROs (16.7%). the prevalence transitions Establishments transitions this and of Rule 2 All only together, 16% of the four types of if Smooth transitions Shifts falling Rough under the Shifts remit together, account for with instantiation utterances, and are this there are more Shifts and Other counted observed and For versions of only the rule 39% do of not correlate much than better with Retains. classifications CP(U 357), 14% of example, entity-coherent transitions (139 out the of frequencies: and ) CB(U the total, )). are cheap in the sense of Strube and Hahn (1999) (i.e., These n − 1 predicts findings n concerning in the vanilla is not instantiation Our should not, major however, conclude the theory general verified. finding lead is us to that matter: is, it is second in way that parameters do that possible to set the parameters such a as to Poesio et al. Centering: A Parametric Theory make all three with claims verified in a statistical of sense. However, because Strong when percentage violations, the parameters whose C1 is the claim most the largest one is find instantiation in which setting matters the controlling trying to an CF all claims in which definition and realization. Considering are satisfied are those utterance a center an contains an associative reference to that center is as realized ficient utterance Strong C1 identifying with suf- finite for to be verified; IS), also C1 is a strong utterances well With sentences instead clauses has positive effect. the resulting instantiations (IF of Strong We as as the two “basic” versions of Rule 1. and verified, Rule also 1 found, however, Rule 2, on that other: there is The a trade-off between Strong C1, on one side, and and mentioned, while the changes to the Rule (see, Table reducing 13). Identifying the violations of Strong utterance with C1, and of 1 just 2 increase realization parameters Rule or those and indirect e.g., in utterances sentences, (to of realization, 1—up results statistically significant increases a lesser in extent) Figures the number allowing 2 of 4)—although violations Rule Rule 1 (GJW 95) to a total of 7.4% Rule 1 in (GJW the IS 83) instantiation so robust (see and in instantiations. and are even these 45 These are that they still verified, impact on Rule changes only even greater 2 (BFP), to the and parameters have which utterance realization weakly verified with an vanilla instantiation. find With more the IF more we many RSH and IS is the instantiations and grammatical-function RET CON ranking, (i.e., without Establishments); than indeed, SSH and As with GFTHERELIN counting in many instantiation than “pure” RET the second most the IS instantiations a Rule 2 only (BFP) if is ranking, verified with instantiations are only common is at the .05 transition. IF result, pronouns level, of and CFs. with On with second-person (45%) CON, instantiations RET, SSH, a or much are counted as realizations the positive side, these RSH, greater percentage 16% of utterances These are classified as either making and results strengthened one a further as EST. can be further proposed by Strube last Hahn change (1999) to the eters: ranking function by and instead param- GFTHERELIN. adopting With the .05 instantiation, This is Rule 2 (BFP) is STRUBE-HAHN verified at the .01 of only this level, rather than no at the on level. of Strong CON C1 (obviously) because although or R1 (more the of SSH surprisingly), RSH. it ranking func- tion has effect find RET does becoming and some the becoming Even result in some we the still Rule more 2 (BFP) RET CON and more .01 RSH SSH, though than than with these cause be verified at the level the IS instantiation. changes are enough to and Hahn’s own to instantiations of Rule 2 still isn’t but this version of Strube instantiation any of version the we verified, In other words, with the rule the IS is verified evaluated. or not by .01 STRUBE-HAHN ranking, all three claims of IF and the theory are verified at the The in final level. observation concerning parameter impact settings on is that issues not widely the the theory’s claims in dis- cussed centering literature had a greater of the issues, choice of ranking function of or our experiments of than parameters Many such as the definition pronouns previous utterance. of these such empty in Considering have to with as the treatment and do the general pronouns issue of second-person included categories, CF which entities should of be the is list. make Strong second-person C1 we to be realiza- tions discourse entities enough to satisfied; also found that Computational Linguistics Volume 30, Number 3 a number of extensions to nonfinite the definition of utterance, much worse such as the inclusion of rela- tive clauses and nonfinite clauses, were led to include results unless reduced relative clauses and in which clauses were taken to traces linking these clauses to the clause they embedded. 5.1.2 Minimizing 3 we Violations Should Not Be minimizing the Overriding Goal. We that think violations should only said in Sec- tion into in in don’t that of how to set parameters. interaction Some be the factor taken account deciding of violations with are best accepted other and explained (See below.) terms the centering preferences preferences. Special or care is needed when of alternative definitions In are supported of by cross-linguistic evidence we find by the results psychological studies. the case ranking, although linear-order didn’t any significant English, differences one between other should keep in grammatical-function mind ranking and ranking for more that ones. such (2000), for languages, in especially Hindi free-order Prasad differences and Strube have been found linear-order for example, found is that the Strube difference between Hahn (1999) grammatical-function and ranking significant, and information and found in German. significant differences versely, between grammatical function and of structure the for slight advantage STRUBE-HAHN Con- over before taking evidence a one needs our ranking ies with grammatical-function ranking as conclusive, to indicating experiments reconciling these results with supplement numerous stud- psychological (among others, important role results the Hudson, Tanenhaus, played by Dell grammatical [1986]; Gordon, function, Grosz, especially Gillion subjecthood Brennan [1995]). Information and structure not and has been found to be appropriate [1993]; including 2002). Greek, Hindi, and Turkish also (Turan 1998; Prasad we considerations apply to the definition of and Strube 2000; for languages Miltsakaki Similar of previous utterance, since saw that a considerable of (Cooreman at when amount psychological evidence supports as least Sanford the 1996; syntactically Pearson, embedded Stevenson, clause Poesio is treating adjuncts embedded, at the 2000). end the In sentence of and of our indicate and with the case the definition finite utterance, results much that identifying sentences, leads to results more utter- ances with rather than clauses, claimed discourses entity-coherent. While consistent the preference number for of to be of this result be for and for different applications (e.g., is likely to useful text we a reasons types planners), believe that further of be about when empirical and theoretical the is work updated. For is reached local focus one needed before conclusions can thing, 1988)–view example, rhetorical structures theory (Mann most analyses Thompson discourse structure—for one the basic unit of discourse in written text. And and clauses as much more can hardly find prosodic any complete of sentences; in in spoken is dialogue this case, the update unit likely to be a phrase some sort. 5.2 The Claims of Centering, Revisited 5.2.1 Centering, Pronominalization, pronominalization—at and Salience. One clear result of this work is that centering’s of Rule 1 proposed claims about in Grosz, Joshi, Weinstein least, (1995, those 1983)–are expressed by robust. the versions (GJW 95) Rule 1 (GJW 83) and instantiations verified we with instantiations, very Rule and all parameter and in 1 way: In are convincing considered, the percentage of a very of Rule 1 (GJW 95) never the one 8% of the in total mind number of violations On the other exceeds utterances. make weak hand, should pronominalization. keep All that these Rule two 1 (GJW versions pronominalize, pronominalize 95) of Rule very about that says is 1 we claims we CB. This that if decide to problem then should of pronominalization the or, more formulation doesn’t address the real for a theory generally, Poesio et al. Centering: A Parametric Theory of NP form decision, pronoun which (Henschel, is to Cheng, decide when Poesio a discourse 2000). And our entity should be realized as a Rule 1 to and “pronominalize results CB,” also which indicate that simply strengthening of the form proposals in Gordon, Grosz, the Gillion can (1993), be seen as a generalization idea: the a Between 50% (with f) and and 60% (with u s) of would be very bad u = = CB not pronoun, conversely, 30% mentions are using a and between and 40% of of the realized personal of pronouns in are not which realizations mismatch of the CB. Examples CB like (12) illustrate situation the and pronominalization one type occur: By sufficiently mentioned a often between having in a discourse, a discourse entity may may been CB. salient (at These the global level) to justify pronominalization even when become not observations whether is the CB decision pronominalize it is the 46 only on suggest that must to a entity the but involve number does not depend discourse a of further constraints and preferences. 47 5.2.2 CT as a Theory of of Coherence: Constraint 1. Another result of 1 this work Rule is 2— that the validity on centering’s of claims about local much coherence—Constraint the parameters than is and depends choice pronominalization. the Strong to a C1 greater not extent the case claims does hold for instan- for the about it instantiation in which the tiation, hold for the implicit vanilla although instantiations of does in which is any treated identified as an with indirect anaphoric component bridging references realization most are (between parameter 25% there are many sentences. more But and for many utterances instantiations, even under exceptions to Strong the C1 favorable instantiations 20% which and of worse the total Rule number 1 (7–8%). of utterances) While than weak we find even of with the with only are instantiation is at most for one most the version C1, requiring that there salient entity per utterance, does with hold even CB the vanilla process, and does capture of the claim that utterances unique central aspect centering since Joshi and Kuhn a (1979) Joshi are easier Weinstein to what (1981), a and and it says nothing about entity coherence’s being Further ensures local Karamanis on coherence. light (2003), entity which coherence is shed by when recent work on text ways planning, of particu- larly by from propositions suggests that all alternative a the expressed by texts such those we extracting text plan ordering in as are the texts tends to be in are studying with considered, the actual found with most greater centering’s about coherence than of its agreement After preferences our the propositions entity 48 expressed by texts in the museum alternatives. extracting optimal Karamanis subdomain minimizing that although the sequence actually found (with in of corpus, is not determined such texts instantiation as far 70% considers, more as than 50% of the violations to entity coherence the utterances violate Strong C1), approx- the imately he If must we of the alternative orderings other ways of in our introduce even more violations. accept that the texts corpus are coherent, these from results what suggest we that there be achieving An local obvious coherence, apart have been calling here “entity coherence”. candidate for an additional, or Computational Linguistics Volume 30, Number 3 alternative, coherence-inducing needs device are rhetorical relations. Indeed, the claim that “entity” coherence Kintsch to be supplemented Dijk (1978) by “relational” Hobbs (1979). coherence This can is be found as far back as of many our and van With and in the u=f which instantiations, we find in view supported by an analysis data. the not pharmaceutical subdomain examples successive is utterances do indicated mention the same entities, in (23): but the connection between clauses explicitly by connectives, as (23) (u1) This Product A. leaflet is a summary of the important information about

(u2) If you have any questions or are not sure about anything to do with your treatment,

(u3) ask your doctor or your pharmacist. A more of complex case CFs are utterances in the museum subdomain that do not refer previous they generic statements about the class of to any objects the of which the object because express will under discussion illustrated is an instance, or object. utterances a point that then be by specific In (24), that (u2) make background generic of a gives concerning the decoration a cabinet: (24) (u1) medallion On the drawer of above Louis the XIV. door, (u2) gilt-bronze In Dutch military Wars trophies of 1672–1678, flank a France portrait the fought simultaneously (u3) against This the Dutch, Spanish, and armies, all. cabinet celebrates the Treaty Imperial Nijmegen, defeating which them war. of concluded the While the analysis of such cases in in terms of of rhetorical relations is more complex, it seems clear or propositions to us that an is analysis more terms underlying one semantic in connections of between events While notion it is that some of perspicuous than true these violations could be fixed terms entity coherence. of in (24) we might by adopting a broader bridging wider reference–for notion of example, proven treat France as difficult a bridge to Louis identify XIV—this in way. bridging reference has to be very to Now, a reliable one given that in an RST-style analysis, every discourse one might unit is wonder connected least link another discourse whether by at rhetorical “entity is to Knott still needed at least coherence” once unit, (2001) “relational in RST, coherence” is introduced. et that complete connectivity is However, al. introducing convincingly argue by “Elaboration” that, when usually achieved out relations such as notion of looked This at turn attempts to entity coherence. work closely, really to be is capture a coming symmetrical to our own: on rhetorical relations is not sufficient, to a position of that a is purely (Knott and a separate theory entity coherence necessary rela- tional account et al. 2001). 49 5.2.3 Topic Continuity: Rule objects, Rule 2. Rule 2–stating a preference not just to keep same irrespective but to preserve of its their relative ranking—also of instantiation. seems much talking about robust the 1, less than formulation and the Poesio et al. Centering: A Parametric Theory As already noted, one of the most interesting of observations about this it aspect concerns utterances used to formalize (at of the theory the of classification With much least instantiations in the we earlier versions of the theory). neither most pretty all common transitions were parameter which CB), the NULL that tested, two the of transition utterances previously considered only in (between two (1998), we ZERO has (from a an utterance with a CB to one Passonneau and the transition far never been discussed before. Indeed, without), with which as as outside 84% can of see has of utterances are either NULL, ZERO, or EST the vanilla in- stantiation, all whether Rule 2 in almost all its formulations. The and therefore fall finding is the scope question other to be extended to cover such cases or raised whether by this the theory has have to for by components of overall theory of they be accounted an discourse (see below). Three Grosz, Joshi, versions of Rule 2 Weinstein were (1995), tested in some in detail. 50 of The version of Rule 2 from and of CON over formulated terms of RET over sequences and (which for sequences number sequences sequences of stating SHIFT a preference we tested with by counting the of instantiations, sequence pairs), involve even the “best” fewer than one-third suffers of of from the prob- lem that sequence pairs the (BFP), IS Grosz same transition, et STRUBE-HAHN Even in and with al. the instantiation even fewer are only which sequences the transitions consid- ered by 13% of yields the best results of for Rule 2 CON-CON/RET-RET/SH-SH, ranking, only 28% sequence of pairs are and sequence pairs involve the form transitions only considered of by Grosz altogether et al. Keeping we find in mind with that Rule IS 2 (GJW STRUBE-HAHN 95) only to a minority applies number of sequence CON-CON pairs, do (37) that and ranking, RET (35), the which in sequences number of SH-SH slightly (19, exceeds of which the 16 number RSH-RSH). of RET- This with turn exceeds GFTHERELIN the where RET-RET are doesn’t IF if we hold EST a type of CON; ranking, we find no exceeds CON-CON even IS treat as significant difference between the and Rule the of 2 (BFP), setting. formulated (single in terms of single utterances) was transitions, accounts for larger centages data and found to be verified both with per- instantiation the NULL with the “best” with instantiations. most instantiations; However, we we still observed the vanilla of and more a large RET percentage CON more RSH transitions SSH in most instantiations in which also found identified than with and are or than allow for indirect utterances sentences realization. 51 Finally, of Strube and ones Hahn’s isn’t preference for sequences of of instantiations cheap transitions we over se- quences in instantiations expensive we verified ones, all CP of studied one we by any found more the tested; in- deed, meaning expensive transitions than utterance doesn’t predict the CB cheap next. that the generally of the These mixed results are in line with those of psychological experiments, far evidence supporting the claim that, say, Continuations which so haven’t process found clear are 1993) easier to than Shifts, let alone Retains (Gordon, Grosz, and Gillion Computational Linguistics Volume 30, Number 3

While proposing our modifications of broad centering is beyond the scope of worthy this of article, we believe results do have theoretical consequences further explo- ration. 5.3.1 Clarification of the Claims ways of and Identification of Further Parameters. setting parameters already discussed in Apart from comparing our work different identifying more the fundamental need goal of made clarifying more the claims of the literature, had the of be precise. Our centering theory by aspects that to in definitions or only of the concepts used in study raised centering not a number questions about the in previously examined Many of of the literature discussed theory. these have to with passing. One questions is do realization, of one of the least studied as- pects pronouns. the Our indicated such question if PRO2s the status if that are not entities realized as second-person we results R1-pronouns, we find many more considered of realizations Strong C1 of CF, or 1, treat them (although as violations Constraint claims are still verified). We and Rule respectively 1 both Rule 1 on whether also saw that the results concerning non-finite VPs were and depended R1-pronouns. More to contain generally, traces we and whether reduced identified or not relative clauses need these traces were and assumed assumed nition of “R1-pronoun”: to be is, whether we include the in for a clear defi- implicit that of should traces relative pronouns clauses, the “pronouns” anaphoric elements which Rule bridging among to 1 references, of, indicate applies. This and demonstrative the we our question isn’t mentioned literature that, for example, implicit in the know of yet results or pronouns treating R1-pronouns the anaphoric is elements idea. bridging references second-person as a Some very bad of One the issues raised is by this study are of relevant only for certain parameter in- stantiations. example the specification whether grammatical NPs in function there-clauses ranking beyond the simplest cases: or for objects example, (our postcopular or nominal should as suggest the former) how mod- be treated ifiers subjects which should be ranked identified (we results treated with them finite as adjuncts). is what An issue for instantiations in when utterances are finite is in middle clauses of the finite previous utterance is an embedded in clause the from another at the following example, the Guardian : clause, rather than the end, as (25) But Hutchinson, who appointed Ranieri last season , today said that he 5.3.2 man, Separating Pollard Entity (1987) Coherence more from recently, CB Uniqueness. Beaver (2004) Starting Kibble with (2001) Brennan, Fried- and and, of original and proposed there have been We attempts work to “unpack” some our the preferences of by centering. it one would feel this has greatly helped to Constraint understanding 1 into the theory and believe be useful unpack two separate claims, as well: that similarly The about first uniqueness of of the (both CB, and one of) about Constraint entity coherence. We will function versions : 1 is to claim that the CB is unique. call this claim CB uniqueness CB Uniqueness: Utterances have at most one CB. We well: have argued throughout the article that Strong Constraint not occur 1 has a second function as to express a preference for utterances that do at the beginning of a Poesio et al. Centering: A Parametric Theory segment to mention Kibble at (2000) least one of the Karamanis objects (2001), included we will in the previous utterance. of Fol- lowing 1 and : call this second half Constraint (entity) continuity (Entity) Continuity: CF(U i 1 ) CF(U i ) = − Weak C1 is CB uniqueness, whereas Strong C1 is CB uniqueness plus continuity. 5.3.3 A Hybrid View of Coherence. of One need clear conclusion suggested by our induce coherence to be supplemented by accounts of results other is that entity-based accounts factors include that in coherence at the local level. of The most may direct way to do this would be to one continuity a in longer order list factors that link an utterance to its previous one of and claim must that The for an utterance to be would “locally coherent,” at least these factors be present. resulting claim take a form along the following lines: Hybrid Continuity For every utterance U i , at least one of the following must hold: 1. CF(U i 1 ) CF(U i ) = − ; 2. There is a rhetorical relation RR such that RR(U i 1 ,U i ). 52 − 3. U i − 1 and of Kameyama, U i are temporally Passonneau, coherent in Poesio the sense, (1993). for example, and 4. ... (other factors) A more sensible would approach, more especially as we don’t yet know of all the factors not be explicit about the scope centering theory, affecting coherence, it be to as a of account of coherence,” but only viewing comprehensive “local In other words, as an (Entity) entity coherence to local coherence. we account of the contribution A natural Continuity only one could view way as among would the preferences include Entity holding Continuity at the discourse level. to formalize proposed this Beaver, be to which would include among a set of constraints like those by also have to further constraints specifying preferences for rhetorical and temporal coherence. 5.3.4 CB Uniqueness. Weak We C1, saw or in CB Section 4 that it’s All fairly is easy needed to fix is the problem of utterances’ violating on uniqueness: it that to which to in strenghten the requirements the ranking function and require be total, turn can be easily done by adding a Before disambiguation factor however, to ranking we functions that like function. doing this, should whether aren’t total grammatical we from finding CB ask will this is the with conclusion should draw if instead the we that uniqueness be violated more partial one ranking CB. functions—or should allow for utterances to have When than multi-CB one utterances such ( as (10) are considered, (Branicki), entity neither occupies the corner cupboard ) is more it is not immediately ob- vious that discourse salient than the in one particularly salient position either in other especially utterance (u227) since or a (u229). Notice the previous mentioned the current one of is also that ( both entities have been before; and furthermore, them animate Branicki ), the Computational Linguistics Volume 30, Number 3 other inanimate ( the cupboard Sidner ). In these (1979) respects, these examples are reminiscent imate led to argue for two foci—sentences with one of the examples that (typically in inanimate one (typically in an- entity agent position) and an yesterday. like Although sold book from for 10 or theme position), Mortimer the cents Mary took Gordon, Grosz, a nickel from her Gillion toy bank when the results studies such as only one and (1993) suggest that we not two animate of entities are considered, materials tends to show RNP effects, Sidner. are aware any experiment testing like those discussed by The hypothesis number that of topicality is not restricted it to one is entity most per utterance has been advanced with work by a of Givon (1983). researchers, Within although perhaps clearly associated we the the centering Gundel literature, (1998) abandoning Strube has Gordon been suggested Hendrick by (1999), and, more the in call “CB uniqueness” (1998), Tetreault radically, claim that work which such whole as notion of CB is and and (2001) in As the seen in Section 2, abandoned. Inference the primary in motivation monadic is for CB uniqueness is complexity- theoretic (Joshi arguments: Kuhn 1979; Joshi Weinstein logics 1981). less Grosz expensive than in normal logics and CB is and in (3), and colleagues’s lin- guistic pronominalize evidence for uniqueness ( Susan, contrasts in like those pronominalize is a more showing that certain that example) serious problem failing to entities others Name ( Betsy). (Gordon, This claim Grosz, is than failing to Repeated Penalty further supported Gillion 1993). by the However, evidence concerning RNP is the mentions observed only in a provided of and the subset the cases Constraint that would 3, be in considered as to by and the example we CB according are would (10), neither the definition discussing RNP Branicki nor the cupboard Gordon occur in In (u229) other in words, a position (some) that be subject to (u229) by Grosz effects et al. in according support of to CB et al. evidence used in (10) CB. This uniqueness is cannot be with used to argue that problem has a single evidence (10): Instead also consistent to by examples like of a different solution CB the raised Givon ranking (1983) function Gundel to be (1998). total, In one attempting to preserve uniqueness by requiring in the could abandon need of RNP and More both cases, we would CB uniqueness, as suggested on a separate issue. theoretical account effects. empirical evidence is needed this 53 5.3.5 Variety. The important third conclusion suggested in by our production results is that This as a principle discourse as maintaining ensuring variety seems to be ence. is R1-pronouns. first is of all, by the fact that CBs only slightly over half of CBs coher- suggested, more or It are realized as also the case that are hardly ever is continued for than two three utterances; of that NP the same in discourse row (even entity with pronouns, very unlikely to only be realized 58 pronoun-pronoun using the same type twice involve of a we have sequences—26% the total); In and we that wonder two-thirds observed different Gordon transitions. might fact, not instance whether of all transition sequences more repeated name two that the penalty by et al. be an of this general phenomenon. Poesio et al. Acknowledgments Special Alistair thanks Knott, to Nikiforos Mark Karamanis, Mitkov, Jon Liberman, Oberlander, members Tim Ruslan other of Rakow, GNOME and the Kees Kibble, Deemter, the Jamie Renate project: van Pearson, Henschel, Donia Rodger wish James and Arnold, to thank Bird, Susan Allen, Scott. Jennifer We also Byron, Steve Herb Clark, Brennan, Jeanette George Donna Gundel, Aravind Ferguson, Miltsakaki, Rashmi Joshi, Len Schubert, Prasad, Eleni Harold Ellen Somers, Prince, Lyn Walker, Joel Tetreault, University and of audiences Pennsylvania, at the ACL 2000, the University of Rochester, CLUK, the University of Wolverhampton and the was The for comments and suggestions. Debbie corpus De presented Jongh, Ben here Donaldson, annotated Marisa by Fraser, Michael Green, Flecha-Garcia, Shane Montague, Camilla Carol with Rennie, and A Claire Thomson, together work, the including authors. substantial of part of this was the creation GNOME, GR/L51126/01. by the EPSRC Massimo project the corpus, supported was of Poesio supported EPSRC during parts this project by Di Eugenio an was advanced fellowship. in Barbara INT 9996195, supported in part Janet part by NATO by NSF grant CRG 9731157. Hitzeman was in grant EPSRC project SOLE, part supported GR/L50341. by the References Alshawi, Language Hiyan. Interpretation 1987. Memory . Cambridge and Context Arnold, University for Jennifer Press, E. 1998. Cambridge. Discourse Patterns . Ph.D. Reference Stanford Form and University, Stanford CA. thesis, Baker, John Collin B. Lowe. F., Charles 1998. The J. Fillmore, Berkeley and FrameNet Proceedings project, of 36th Montreal, , Canada. 86–90. In Beaver, David. 2004. the The ACL optimization pages Linguistics of discourse Philosophy anaphora. , 27(1):3–56. and Brennan, in Susan E. Language 1995. Centering Cognitive attention Processes discourse. , 10:137–167. and Brennan, Charles Susan J. E., Pollard. Marilyn 1987. W. Friedman, and approach ACL, Stanford, pronouns. CA, In Proceedings A centering 25th to 155–162, of June. the Byron, Donna model Amanda pages and of Stent. in 1998. A preliminary In Proceedings of 36th centering , Montreal, dialog. Canada. the ACL Centering: A Parametric Theory Caramazza, Catherine Alfonso, Garvey, Ellen Grober, Comprehension of and Jack Yates. 1977. Journal of Verbal anaphoric Learning pronouns. Behavior , 16:601–609. and Verbal Carletta, Jean. 1996. Assessing Computational The agreement on classification tasks: Linguistics kappa , 22(2):249–254. statistic. Chafe, Wallace. 1976. Givenness, contrastiveness, In C. Li, editor, definiteness, subjects, Subject Topic and topics. Academic Chinchor, Press, New A. York, and . Nancy Beth pages 25–76. 1995. Message and Sundheim. (MUC) of understanding processing. conference Proceedings tests on of discourse Methods the AAAI in Spring In Empirical Symposium Interpretation Discourse Stanford, CA. and Generation , pages 21–26, Clark, Johnson-Laird Herbert H. 1977. P. C. Bridging. Wason, In P. N. Thinking: Readings and in Cognitive Science editors, 411–420, Cambridge University , pages Cooreman, Press, London Ann and Tony New Sanford. York. Focus and 1996. and Research syntactic Paper subordination no. RP-79, in discourse. Cote, University Sharon. of 1998. Edinburgh, Ranking HCRC. forward-looking A. K. Joshi, E. centers. F. Prince, In M. A. Walker, Centering Theory and in Discourse . editors, University Press, Oxford, Oxford, 55–70. chapter 4, Dale, pages Expressions Robert. . 1992. MIT Press, Generating Cambridge, Referring Di Eugenio, Barbara. 1998. Centering in MA. Italian. E. F. Prince, In M. A. Walker, Centering A. K. Joshi, Discourse Oxford editors, University Press, Theory and . in Di Oxford, Eugenio, chapter Barbara, 7, pages Johanna 115–138. Massimo Paolucci. 1997. D. Learning Moore, and features Proceedings that of predict 35th cue usage. , Madrid. In Fox, Barbara A. the 1987. Discourse ACL . Cambridge University Structure Press, and Anaphora Gernsbacher, Cambridge. David Hargreaves. Morton A. 1988. and first participants: The Accessing sentence mention. Journal of Memory advantage of Language , 27:699–717. and Giouli, Paraskevi. 1996. in Topic chaining and discourse structure Master’s task-oriented dialogues. Department, University thesis, of Edinburgh. Linguistics Givon, in Discourse: Talmy, editor. Quantitative 1983. Topic Cross-Language Continuity Study. Benjamins, A Philadelphia. Amsterdam and Computational Linguistics Gordon, Laura Peter A. C., Barbara Gillion. 1993. J. Grosz, Pronouns, and of in names, and the centering Cognitive attention discourse. Gordon, Peter C. Science Randall , 17:311–348. 1999. The and Hendrick. of representation in and processing Science, coreference 22:389–424. discourse. Cognitive Gordon, Kerry Peter C., Ledoux, Randall Hendrick, Processing of and Chin L. Yang. 1999. of An reference and of the structure language: Language analysis Cognitive complex noun 14(4):353–379. phrases. and Processes , Grosz, Use Barbara of Focus J. in 1977. The Dialogue Representation Understanding . and Ph.D. Stanford, thesis, Stanford CA. University, Grosz, Scott Barbara Weinstein. J., Aravind 1983. Providing K. Joshi, and of noun a in unified account Cambridge, In definite MA, Proceedings phrases discourse. 44–50. of ACL-83 , Grosz, Scott Barbara J., Aravind pages Weinstein. 1986. Towards K. Joshi, and of a computational interpretation. Unpublished theory discourse Grosz, Barbara J., Aravind K. Joshi, manuscript. Scott Weinstein. 1995. Centering: A and framework of for modeling Computational the local coherence Linguistics , 21(2):202–225. discourse. Grosz, 1986. Barbara Attention, J. and intention, Candace L. Sidner. of Computational and the Linguistics structure , 12(3):175–204. discourse. Gundel, Jeanette K. 1998. Centering Towards theory and the givenness In M. A. hierarchy: Walker, A. K. Joshi, a synthesis. and in Discourse E. F. Prince, . Oxford editors, University Centering Press, Theory Gundel, Oxford, Jeanette chapter K., 10, Nancy pages Hedberg, 183–198. Ron Zacharski. 1993. Cognitive and of status in and the form Language referring Hawkins, A. 1978. , expressions discourse. John 69(2):274–307. Indefiniteness . Croom Helm, Definiteness London. and Heim, Indefinite Irene. 1982. Noun The Phrases Semantics . Ph.D. of Definite and University of Massachusetts Amherst. thesis, Henschel, Renate, Hua Cheng, at Massimo Poesio. 2000. Pronominalization and revisited. Saarbruecken, In Proceedings August. of 18th COLING , Hitzeman, Chris Mellish, Janet, Alan Jon Black, Oberlander. Paul Taylor, On of and 1998. the use automatically information in generated discourse-level a concept-to-speech Proceedings of International synthesis system. Conference In the on Volume 30, Number 3 Spoken Language Sydney, Processing Australia. (ICSLP98), Hitzeman, Janet Long-distance and Massimo Poesio. pronominalisation 1998. In Proceedings of and global focus. ACL/COLING Montreal. , vol. 1, pages 550–556, Hobbs, Jerry R. 1979. Coherence Cognitive Science , and Hudson, coreference. Susan B., Michael K. 3:67–90. Gary S. Dell. 1986. The Tanenhaus, and on effect of the discourse center In Proceedings the local of coherence Eight of a discourse. Society, Meeting of the Cognitive the Annual Amherst, MA, Science Hudson-D’Zmura, Susan pages 96–101. Tanenhaus. 1998. Assigning and Michael K. pronouns: The antecedents to ambiguous of role of the center attention In M. A. as Walker, the default assignment. E. F. Prince, Centering A. K. Joshi, and in Discourse . Oxford editors, University Press, Theory Hurewitz, Oxford, pages Felicia. 199–226. 1998. A In quantitative M. A. Walker, look at A. discourse Centering K. Joshi, coherence. Theory and in E. Discourse F. Prince, . editors, University Press, Oxford, Oxford Joshi, Aravind K. Steve pages Kuhn. 273–291. Centered The and role of 1979. logic: in natural entity centered sentence representation language IJCAI, Tokyo, inferencing. 435–439. In Proceedings of the Joshi, Aravind K. pages Control of inference: and Scott Role Weinstein. of 1981. of some In aspects Proceedings discourse of structure–centering. 385–387. the IJCAI , Vancouver, CA, Kameyama, pages The Case of Megumi. Japanese . 1985. Ph.D. Zero Anaphora: University, Stanford, CA. thesis, Stanford Kameyama, property-sharing Megumi. 1986. A In Proceedings of constraint in , New centering. 200–206. the ACL-86 York, Kameyama, pages Megumi. A study. 1998. In Intra-sentential centering: A. K. Joshi, case E. F. Prince, M. A. Walker, Centering Theory and in Discourse . editors, University Press, Oxford, Oxford 89–112. chapter 6, Kameyama, pages Massimo Megumi, Poesio. Rebecca 1993. Temporal Passonneau, and centering. 70–77, In Columbus, Proceedings OH. of the 31st ACL , Kamp, pages Discourse Hans and Logic Uwe . Reidel, Reyle. Dordrecht. 1993. From Karamanis, to Nikiforos. 2001. Fourth CLUK . University In Exploring entity-based coherence. Proceedings of Sheffield, of the pages 18–26. Poesio et al. Karamanis, Nikiforos. Descriptive Text 2003. Entity Structuring . Coherence for Informatics, University Ph.D. thesis, Edinburgh. of Karttunen, In J. Lauri. 1976. McCawley, Discourse Semantics editor, from Syntax referents. 7—Notes and Underground . Academic the Linguistic New York, Press, Kibble, Rodger. pages 363–385. 2 of 2000. A reformulation Technical of rule ITRI, centering University theory. Kibble, report, Rodger. 2001. A of Brighton. Rule 2 of reformulation of Linguistics centering , 27(4):579–587. theory. Computational Kibble, Rodger integrated and Richard Power. 2000. An pronominalization. framework for In text Proceedings planning and INLG, Mitzpe Ramon, Israel, June. of Kintsch, Towards Walter model and of Teun van Dijk. 1978. a Psychological and , production. discourse comprehension Knott, Alistair, Review Jon Oberlander, 85:363–394. Mick Beyond O’Donnell, and The Chris interaction Mellish. of 2001. elaboration: relations T. Sanders, and J. Schilperoord, focus in coherent text. In W. Spooren, Text Representation: and Linguistic editors, Benjamins, and Amsterdam Psycholinguistic Philadelphia, Aspects . John 181-196. and Lappin, pages An Shalom and pronominal Herbert J. Leass. 1994. algorithm Computational for Linguistics anaphora resolution. 20(4):535–562. , Lascarides, Temporal Alex interpretation, and Nick Asher. 1993. discourse relations Linguistics and commonsense William C. Philosophy Sandra , 16(5):437–493. entailment. Mann, and Thompson. 1988. and Rhetorical A. Towards structure theory: organization. a functional Text, 8(3):243–281. theory of Marcu, text manually Daniel. 1999. Instructions for annotating of Unpublished the discourse structures manuscript, USC/ISI. texts. McKeown, Kathy R. 1985. Discourse strategies for generating Intelligence natural-language Miltsakaki, text. Artificial Eleni. 1999. Locating , 27(1):1–41. processing. In Proceedings topics of CLIN in text Utrecht, 127–138. , Miltsakaki, pages Eleni. of 2002. Towards an aposynthesis intrasentential topic continuity Computational and Linguistics , 28(3):319–355. anaphora. Moser, Toward Megan and Johanna of D. Moore. of 1996. a synthesis two accounts Centering: A Parametric Theory discourse Linguistics structure. , Computational Oberlander, 22(3):409–419. Alistair Jon, Knott, Mick O’Donnell, Conversation and in Chris Mellish. museum: 1998. Experiments in the intelligent dynamic hypermedia of labelling explorer. with the Review Hypermedia New 4:11–32. and Multimedia , Oberlander, Entity Jon and Massimo Poesio. 2002. coherence investigation. and relational Paper coherence: a corpus-based presented Topics in at the Berlin Discourse, workshop September. on Passonneau, Rebecca J. of 1993. Getting Bates R. M. Weischedel, attention. In and keeping M. the center Challenges and in Natural Language Processing editors, Cambridge University Press, Cambridge, . Passonneau, chapter 7, pages Rebecca 179–227. J. 1997. Instructions for applying discourse multiple reference annotation (DRAMA). for Unpublished applications Passonneau, Rebecca J. 1998. manuscript. with Interaction of of discourse structure noun explicitness discourse M. A. Walker, anaphoric A. K. Joshi, phrases. E. F. In Prince, Discourse . Oxford Centering University Theory and editors, Press, in Passonneau, Oxford, chapter Rebecca 17, J. pages Diane 327–358. 1993. Feasibility of and Litman. In Proceedings automated Meeting Columbus, of discourse segmentation. of , OH, 31st Annual 148–155. the ACL pages Pearson, Massimo Jamie, Poesio. Rosemary 2000. Pronoun Stevenson, and in In Proceedings resolution AMLAP, complex Leiden. sentences. of Pearson, Massimo Jamie, Poesio. Rosemary 2001. The Stevenson, of and animacy, role, effects on thematic and of surface position In the M. focusing Poesio, editor, entities of SEMPRO , University of Proceedings in discourse. Edinburgh. the First Poesio, Massimo. 2000. Annotating a corpus to develop and evaluate Issues discourse entity realization algorithms: 2nd LREC results. 211–218, In Proceedings and preliminary , Athens, May. of the Poesio, Massimo. pages 2003. Associative descriptions EACL Workshop and salience. on Computational In Proceedings of the Poesio, Treatments Massimo. of Anaphora 2004. The , Budapest. MATE/GNOME scheme In for , revisited. May. Proceedings anaphoric annotation, SIGDIAL Boston, of Computational Linguistics Poesio, Massimo, Laurent Florence Romary. 1999. Bruneseaux, The MATE and meta-scheme in multiple for coreference In M. in dialogues editor, languages. Proceedings of Walker, Standards Tools the ACL Workshop Discourse Tagging on and 65–74. for , Poesio, pages 2001. Massimo Discourse and Barbara Di Eugenio. accessibility. In Ivana structure Kruijff-Korbayov and anaphoric Mark Steedman, Proceedings ´ a and of ESSLLI 2001 Workshop editors, Structure, the Discourse on Structure Information Semantics , Helsinki. and Discourse Poesio, Massimo Natalia Nygren-Modjeska. and THIS-NP A 2003. The investigation. hypothesis: In Proceedings corpus-based Lisbon. of DAARC , Poesio, Forthcoming. Massimo Empirical Evidence Salience: and Rosemary . Theoretical Stevenson. Cambridge University Models and Poesio, Press, Massimo Cambridge and Renata New Vieira. York. investigation and of 1998. A corpus-based Computational definite description 24(2):183–216. use. Linguistics , Prasad, Discourse Rashmi and Michael pronoun Strube. 2000. salience Papers in Linguistics in Hindi. and resolution , In Penn 6, Working of Pennsylvania, Philadelphia, volume University Prince, pages Ellen 189–208. F. information. 1981. Toward In P. a taxonomy Cole, editor, of given-new Radical York, Pragmatics 223–256. . Academic Press, New Prince, Subjects, Ellen pages F. 1992. The ZPG In S. Thompson and information letter: definiteness, status. Discourse Description: and W. Diverse Mann, editors, Analyses Amsterdam of a Fund-Raising Philadelphia, Text . Benjamins, 295–325. and Quirk, pages 1973. Randolph University and Grammar Sidney Greenbaum. Longman, A Harlow, England. of English . Rambow, of Owen. 1993. Pragmatics aspects German. scrambling In Proceedings and topicalization of Workshop in on Centering Theory in Naturally- the Occurring Research in Discourse Cognitive . Institute Science, for Sanford, Philadelphia. 1981. Understanding Anthony J. and Written Simon Language C. Garrod. Wiley, Chichester, England. . Scott, Roger Donia, Evans. Richard 1998. Power, Generation and its own problem. In as a solution Proceedings to of the ninth International Volume 30, Number 3 Workshop on Natural Language Niagara-on-the-Lake, CA. Generation Sgall, Petr. 1967. in Functional sentence perspective Prague Studies in a generative description. 2:203–225. Mathematical Linguistics Sidner, Candace Computational L. 1979. Comprehension Theory in of Towards Definite a English Discourse Anaphora Massachusetts Institute of . Ph.D. thesis, Technology, Siegel, Sidney Cambridge. Nonparametric and N. John Statistics Castellan. 1988. Sciences, 2nd for the Behavioral Stevenson, Rosemary edition McGraw-Hill, J., Rosalind A. Boston. Crawley, Thematic and David roles, Kleinman. 1994. of focus, and Language the representation Cognitive Processes events. , 9:519–548. and Stevenson, Jon Oberlander, Rosemary, Alistair Scott McDonald. Knott, 2000. Interpreting and Interactions pronouns and connectives: roles between focusing, thematic Language Cognitive and coherence 15, relations. 225–262. and Processes , Strube, pages Michael. 1998. Never In look Proceedings back: An alternative COLING-ACL to centering. , 1251–1257, Montreal. of Strube, Functional Michael pages centering–Grounding and Udo Hahn. 1999. referential Computational coherence in information structure. Suri, 25(3):309–344. Linguistics , RAFT/RAPR Linda Z. and Kathleen F. McCoy. A 1994. and centering: comparison processing and discussion of problems related Computational to Linguistics complex , 20(2):301–317. sentences. Tetreault, Joel of R. 2001. A corpus-based evaluation Computational centering and Linguistics pronoun resolution. 27(4):507–520. , Turan, Umit. in Turkish: 1998. Ranking Universal forward-looking centers A. Joshi, properties. E. F. In and language-specific Walker, K. Prince, M. A. Centering Theory and editors, Oxford University Press, in Oxford, Discourse . Vallduvi, chapter Enric. 8, pages 1990. 139–160. Component . Ph.D. The Informational Pennsylvania, Philadelphia. thesis, University of van On Deemter, Kees Coreference and Rodger in Kibble. MUC 2000. coreferring: Computational and related Linguistics annotation , 26(4):629–637. schemes. Walker, Marilyn processing A. 1989. Evaluating discourse Proceedings , algorithms. 251–261, In Vancouver, ACL-89 British Columbia, pages Canad, June. , , Poesio et al. Walker, Marilyn A. 1993. In Initial contexts Proceedings of and shifting centers. Workshop on Centering , the Pennsylvania, University Philadelphia. of Walker, Marilyn Sharon Cote. A., Masayo 1994. Iida, Japanese and process of discourse and Computational the centering. Walker, Marilyn Linguistics A., , Aravind 20(2):193–232. Ellen F. Prince. 1998a. K. Joshi, Centering in and naturally occurring discourse: An Centering: A Parametric Theory overview. E. F. In M. Prince, A. Walker, A. K. Joshi, and in Discourse . editors, Oxford Centering University Theory Oxford, 1, 1–28. Press, Walker, chapter Marilyn A., Aravind pages Ellen F. Prince, K. Joshi, in Discourse . 1998b. and Theory editors. Oxford Centering Press, Oxford. University Webber, Bonnie L. 1978. A to anaphora. Report formal approach BBN, discourse Cambridge, MA. no. 3761,