Ngày đăng: 31/03/2014, 04:20
Specifying the Parameters of Centering Theory: a Corpus-BasedEvaluation using Text from Application-Oriented DomainsM. Poesio, H. Cheng, R. Henschel, J. Hitzeman, R. Kibble, and R. StevensonUniversity of Edinburgh, ICCS and HCRC,poesio,huac,henschel @cogsci.ed.ac.ukThe MITRE Corporation, email@example.comUniversity of Brighton, ITRI, Rodger.Kibble@itri.bton.ac.ukUniversity of Durham, Psychology and HCRC, Rosemary.Stevenson@durham.ac.ukAbstractThe deﬁnitions of the basic concepts,rules, and constraints of centering the-ory involve underspeciﬁed notions suchas ‘previous utterance’, ‘realization’,and ‘ranking’. We attempted to ﬁnd thebest way of deﬁning each such notionamong those that can be annotated reli-ably, and using a corpus of texts in twodomains of practical interest. Our mainresult is that trying to reduce the num-ber of utterances without a backward-looking center (CB) results in an in-creased number of cases in which somediscourse entity, but not theCB,getspronominalized, and viceversa.1 MOTIVATIONCentering Theory (Grosz et al., 1995; Walker etal., 1998b) is best characterized as a ‘parametric’theory: its key deﬁnitions and claims involve no-tions such as ‘utterance’, ‘realization’, and ‘rank-ing’ which are not completely speciﬁed; their pre-cise deﬁnition is left as a matter for empirical re-search, and may vary from language to language.A ﬁrst goal of the work presented in this paperwas to ﬁnd which way of specifying these param-eters, among the many proposed in the literature,would make the claims of centering theory mostaccurate as predictors of coherence and pronomi-nalization for English. We did this by annotatinga corpus of English texts with the sort of informa-tion required to implement some of the most pop-ular variants of centering theory, and using thiscorpus to automatically check two central claimsof the theory, the claim that all utterances have abackward looking center (CB) (Constraint 1), andthe claim that if any discourse entity is pronomi-nalized, theCB is (Rule 1). In doing this, we triedto make sure we would only use information thatcould be annotated reliably.Our second goal was to evaluate the predic-tions of the theory in domains of interest for realapplications–natural language generation, in ourcase. For this reason, we used texts in two gen-res not yet studied, but of interest to developers ofNLG systems: instructional texts and descriptionsof museum objects to be displayed on Web pages.The paper is organized as follows. We ﬁrst re-view the basic notions of the theory. We then dis-cuss the methods we used: our annotation methodand how the annotation was used. In Section 4 wepresent the results of the study. A discussion ofthese results follows.2 FUNDAMENTALS OF CENTERINGTHEORYCentering theory (Grosz et al., 1995; Walker etal., 1998b) is an ‘object-centered’ theory of textcoherence: it attempts to characterize the textsthat can be considered coherent on the basis ofthe way discourse entities are introduced and dis-cussed.1At the same time, it is also meant tobe a theory of salience: i.e., it attempts to pre-dict which entities will be most salient at anygiven time (which should be useful for a naturallanguage generator, since it is these entities thatare most typically pronominalized (Gundel et al.,1993)).According to the theory, everyUTTERANCE ina spoken dialogue or written text introduces intothe discourse a number ofFORWARD-LOOKINGCENTERS(CFs). CFs correspond more or less1For a discussion of ‘object-centered’ vs. ‘relation-centered’ notions of coherence, see (Stevenson et al., 2000).to discourse entities in the sense of (Karttunen,1976; Webber, 1978; Heim, 1982), and can belinked toCFs introduced by previous or suc-cessive utterances. Forward-looking centers areRANKED, and because of this ranking, some CFsacquire particular prominence. Among them, theso-calledBACKWARD-LOOKING CENTER (CB),deﬁned as follows:Backward Looking Center (CB)CB(U ), theBACKWARD-LOOKING CENTER of utter-ance U, is the highest ranked element ofCF(U ) that is realized in U .Utterance Uis classiﬁed as a CONTINUE ifCB(U )=CB(U )andCB(U )isthemosthighly rankedCF of U ;asaRETAIN if the CBremains the same, but it’s not any longer the mosthighly-rankedCF;andasaSHIFT if CB(U )CB(U ).The main claims of the theory are articulated interms of constraints and rules onCFsandCB.Constraint 1: All utterances of a segment exceptfor the 1st have exactly oneCB.Rule 1: if anyCF is pronominalized, the CB is.Rule 2: (sequences of) continuations are pre-ferred over (sequences of) retains, which arepreferred over (sequences of) shiftsConstraint 1 and Rule 2 express a preference forutterances in a text to talk about the same ob-jects; Rule 1 is the main claim of the theory aboutpronominalization. In this paper we concentrateon Constraint 1 and Rule 1.One of the most unusual features of centeringtheory is that the notions of utterance, previousutterance, ranking, and realization used in the def-initions above are left unspeciﬁed, to be appropri-ately deﬁned on the basis of empirical evidence,and possibly in a different way for each language.As a result, centering theory is best viewed as acluster of theories, each of which speciﬁes theparameters in a different ways: e.g., ranking hasbeen claimed to depend on grammatical function(Kameyama, 1985; Brennan et al., 1987), on the-matic roles (Cote, 1998), and on the discourse sta-tus of theCFs (Strube and Hahn, 1999); there areat least two deﬁnitions of what counts as ‘previ-ous utterance’ (Kameyama, 1998; Suri and Mc-Coy, 1994); and ‘realization’ can be interpretedeither in a strict sense, i.e., by taking aCF to berealized in an utterance only if anNP in that utter-ance denotes thatCF, or in a looser sense, by alsocounting aCF as ‘realized’ if it is referred to in-directly by means of a bridging reference (Clark,1977), i.e., an anaphoric expression that refers toan object which wasn’t mentioned before but issomehow related to an object that already has, asin the vase the handle (see, e.g., the discussionin (Grosz et al., 1995; Walker et al., 1998b)).3 METHODSThe fact that so many basic notions of centeringtheory do not have a completely speciﬁed def-inition makes empirical veriﬁcation of the the-ory rather difﬁcult. Because any attempt at di-rectly annotating a corpus for ‘utterances’ andtheirCBs is bound to force the annotators to adoptsome speciﬁcation of the basic notions of the the-ory, previous studies have tended to study a par-ticular variant of the theory (Di Eugenio, 1998;Kameyama, 1998; Passonneau, 1993; Strube andHahn, 1999; Walker, 1989). A notable exceptionis (Tetreault, 1999), which used an annotated cor-pus to compare the performance of two variantsof centering theory.The work discussed here, like Tetreault’s, is anattempt at using corpora to compare different ver-sions of centering theory, but considering also pa-rameters of centering theory not studied in thisearlier work. In particular, we looked at differentways of deﬁning the notion of utterance, we stud-ied the deﬁnition of realization, and more gener-ally the role of semantic information. We did thisby annotating a corpus with information that hasbeen claimed by one or the other version of cen-tering theory to play a role in the deﬁnitions ofits basic notions - e.g., the grammatical functionof anNP, anaphoric relations (including infor-mation about bridging references) and how sen-tences break up into clauses and subclausal units–and then tried to ﬁnd out the best way of specify-ing these notions automatically, by trying out dif-ferent conﬁgurations of parameters, and countingthe number of violations of the constraints andrules that would result by adopting a particularparameter conﬁguration.The DataThe aim of our project, which is calledGNOME and whose home page is athttp://www.hcrc.ed.ac.uk/gnome,is to developNP generation algorithms whosegenerality is to be veriﬁed by incorporatingthem in two distinct systems: theILEX systemdeveloped at the University of Edinburgh, thatgenerates Web pages describing museum objectson the basis of the perceived status of its user’sknowledge and of the objects she previouslylooked at (Oberlander et al., 1998); and theICONOCLAST system, developed at the Univer-sity of Brighton, that supports the creation ofpatient information leaﬂets (Scott et al., 1998).The corpus we collected includes texts fromboth the domains we are studying. The textsin the museum domain consist of descriptionsof museum objects and brief texts about theartists that produced them; the texts in thepharmaceutical domain are leaﬂets providing thepatients with the legally mandatory informationabout their medicine. The total size of the corpusis of about 6,000 NPs. For this study we usedabout half of each subset, for a total number ofabout 3,000NPs, of which 103 are third personpronouns (72 in the museum domain, 31 in thepharmaceutical domain) and 61 are third-personpossessive pronouns (58 in the museum domain,3 in the pharmaceutical domain).AnnotationPrevious empirical studies of centering theorytypically involved a single annotator annotat-ing her corpus according to her own subjectivejudgment (Passonneau, 1993; Kameyama, 1998;Strube and Hahn, 1999). One of our goals wasto use for our study only information that couldbe annotated reliably (Passonneau and Litman,1993; Carletta, 1996), as we believe this willmake our results easier to replicate. The pricewe paid to achieve replicability is that we could-n’t test all hypotheses proposed in the literature,especially about segmentation and about ranking.We discuss some of the problems in what follows.(The latest version of the annotation manual isavailable from theGNOME project’s home page.)We used eight annotators for the reliability studyand the annotation.Utterances Kameyama (1998) noted that iden-tifying utterances with sentences is problematicin the case of multiclausal sentences: e.g., gram-matical function ranking becomes difﬁcult tomeasure, as there may be more than one sub-ject. She proposed to use all and only tensedclauses instead of sentences as utterance units,and then classiﬁed ﬁnite clauses into (i) utter-ance units that constitute a ’permanent’ updateof the local focus: these include coordinatedclauses and adjuncts) and (ii) utterance units thatresult in updates that are then erased, much asin the way the information provided by subor-dinated discourse segments is erased when theyare popped. Kameyama called theseEMBED-DED utterance units, and proposed that clausesthat serve as verbal complements behave this way.Suri and McCoy (1994) did a study that led themto propose that some types of adjuncts–in particu-lar, clauses headed by after and before–should betreated as ‘embedded’ rather than as ‘permanentupdates’ as suggested by Kameyama; these re-sults were subsequently conﬁrmed by more con-trolled experiments Pearson et al. (2000). Nei-ther Kameyama nor Suri and McCoy discuss par-entheticals; Kameyama only brieﬂy mentions rel-ative clauses, but doesn’t analyze them in detail.In order to evaluate these deﬁnitions of ut-terance (sentences versus ﬁnite clauses), as wellas the different ways of deﬁning ‘previous utter-ance’, we marked up in our corpus what we called(DISCOURSE) UNITS. These include clauses, aswell as other sentence subconstituents which maybe treated as separate utterances, including paren-theticals, preposedPPs, and (the second elementof) coordinatedVPs. The instructions for mark-ing up units were in part derived from (Marcu,1999); for each unit, the following attributes weremarked:utype: whether the unit is a main clause,a relative clause, appositive, a parenthetical,etc.verbed: whether the unit contains a verb ornot.ﬁnite: for verbed units, whether the verb isﬁnite or not.subject: for verbed units, whether they havea full subject, an empty subject (expletive,as in there sentences), or no subject (e.g., forinﬁnitival clauses).The agreement on identifying the boundaries ofunits, using thestatistic discussed in (Carletta,1996), was(for two annotators and 500units); the agreement on features(2 annotatorsand at least 200 units) was follows:Attribute Valueutype .76verbed .9ﬁnite .81subject .86NPs Our instructions for identifying NP mark-ables derive from those proposed in theMATEproject scheme for annotating anaphoric relations(Poesio et al., 1999). We annotated attributes ofNPs which could be used to deﬁne their ranking,including:The NP type, cat (pronoun, proper name,etc.)A few other ‘basic’ syntactic features, num,per,andgen, that could be used to identifycontexts in which the antecedent of a pro-noun could be identiﬁed unambiguously;The grammatical function, gf;ani: whether the object denoted is animateor inanimatedeix: whether the object is a deictic refer-ence or notThe agreement values for these attributes are asfollows:Attribute Valueani .81cat .9deix .81gen .89gf .85num .84per .9one of the features of NPs claimed to affect rank-ing (Sidner, 1979; Cote, 1998) that we haven’tso far been able to annotate because of failureto reach acceptable agreement is thematic roles().Anaphoric information Finally, in order tocompute whether aCF from an utterance was re-alized directly or indirectly in the following ut-terance, we marked up anaphoric relations be-tweenNPs, again using a variant of the MATEscheme. Theories of focusing such as (Sidner,1979; Strube and Hahn, 1999), as well as our ownearly experiments with centering, suggested thatindirect realization can play quite a crucial role inmaintaining theCB; however, previous work, par-ticularly in the context of theMUC initiative, sug-gested that while it’s fairly easy to achieve agree-ment on identity relations, marking up bridgingreferences is quite hard; this was conﬁrmed by,e.g., Poesio and Vieira (1998). As a result we didannotate this type of relations, but to achieve areasonable agreement, and to contain somehowthe annotators’ work, we limited the types of re-lations annotators were supposed to mark up, andwe speciﬁed priorities. Thus, besides identity(IDENT) we only marked up three non-identity(‘bridging’ (Clark, 1977)) relations, and only re-lations between objects. The relations we markup are a subset of those proposed in the ‘extendedrelations’ version of theMATE scheme (Poesio etal., 1999) and include set membership (ELE-MENT), subset (SUBSET), and ‘generalized pos-session’ (POSS), which includes part-of relationsas well as more traditional ownership relations.As expected, we achieved a rather good agree-ment on identity relations. In our most recentanalysis (two annotators looking at the anaphoricrelations between 200 NPs) we observed no realdisagreements; 79.4% of these relations weremarked up by both annotators; 12.8% by onlyone of them; and in 7.7% of the cases, one ofthe annotators marked up a closer antecedent thanthe other. Concerning bridges, limiting the re-lations did limit the disagreements among an-notators (only 4.8% of the relations are actuallymarked differently) but only 22% of bridging ref-erences were marked in the same way by both an-notators; 73.17% of relations are marked by onlyone or the other annotator. So reaching agreementon this information involved several discussionsbetween annotators and more than one pass overthe corpus.Segmentation Segmenting text in a reliablefashion is still an open problem, and in additionthe relation between centering (i.e., local focusshifts) and segmentation (i.e., global focus shifts)is still not clear: some see them as independentaspects of attentional structure, whereas other re-searchers deﬁne centering transitions with respectto segments (see, e.g., the discussion in the intro-duction to (Walker et al., 1998b)). Our prelim-inary experiments at annotating discourse struc-ture didn’t give good results, either. Therefore,we only used the layout structure of the textsas a rough indication of discourse structure. Inthe museum domain, each object description wastreated as a separate segment; in the pharmaceu-tical domain, each subsection of a leaﬂet wastreated as a separate segment. We then identiﬁedby hand those violations of Constraint 1 that ap-peared to be motivated by too broad a segmenta-tion of the text.2Automatic computation of centeringinformationThe annotation thus produced was used to au-tomatically compute utterances according to theparticular conﬁguration of parameters chosen,and then to compute theCFsandtheCB (if any)of each utterance on the basis of the anaphoricinformation and according to the notion of rank-ing speciﬁed. This information was the used toﬁnd violations of Constraint 1 and Rule 1. Thebehavior of the script that computes this informa-tion depends on the following parameters:utterance: whether sentences, ﬁnite clauses, orverbed clauses should be treated as utter-ances.previous utterance: whether adjunct clausesshould be treated Kameyama-style orSuri-style.rank: whether CFs should be ranked accordingto grammatical function or discourse statusin Strube and Hahn’s sense2(Cristea et al., 2000) showed that it is indeed possibleto achieve good agreement on discourse segmentation, butthat it requires intensive training and repeated iterations; weintend to take advantage of acorpus already annotated in thisway in future work.realization: whether only direct realizationshould be counted, or also indirect realiza-tion via bridging references.4 MAIN RESULTSThe principle we used to evaluate the differentconﬁgurations of the theory was that the best def-inition of the parameters was the one that wouldlead to the fewest violations of Constraint 1 andRule 1. We discuss the results for each principle.Constraint 1: All utterances of a segmentexcept for the 1st have precisely one CBOur ﬁrst set of ﬁgures concerns Constraint 1:how many utterances have aCB. This con-straint can be used to evaluate how well cen-tering theory predicts coherence, in the follow-ing sense: assuming that all our texts are co-herent, if centering were the only factor behindcoherence, all utterances should verify this con-straint. The ﬁrst table shows the results obtainedby choosing the conﬁguration that comes clos-est to the one suggested by Kameyama (1998):utterance=ﬁnite, prev=kameyama, rank=gf, real-ization=direct. The ﬁrst column lists the numberof utterances that satisfy Constraint 1; the secondthose that do not satisfy it, but are segment-initial;the third those that do not satisfy it and are notsegment-initial.CB Segment Initial NO CB Total NumberMuseum 132 35 245 412Pharmacy 158 13 198 369Total 290 48 443 791The previous table shows that with this conﬁg-uration of parameters, most utterances do not sat-isfy Constraint 1 in the strict sense even if we takeinto account text segmentation (admittedly, a veryrough one). If we take sentences as utterances,instead of ﬁnite clauses, we get fewer violations,although about 25% of the total number of utter-ances are violations:CB Segment Initial NO CB Total NumberMuseum 120 22 85 227Pharmacy 152 8 51 211Total 272 30 136 438Using Suri and McCoy’s deﬁnition of previousutterance, instead of Kameyama’s (i.e., treatingadjuncts as embedded utterances) leads to a slightimprovement over Kameyama’s proposal but stillnot as good as using sentences:CB Segment Initial NO CB Total NumberMuseum 140 35 237 412Pharmacy 167 14 188 369Total 307 49 425 791What about the ﬁnite clause types not consid-ered by Kameyama or Suri and McCoy? It turnsout that we get better results if we do not treat asutterances relative clauses (which anyway alwayshave aCB, under standard syntactic assumptionsabout the presence of traces referring to the modi-ﬁed noun phrase), parentheticals, clauses that oc-cur in subject position; and if we treat as a singleutterance matrix clauses with empty subjects andtheir complements (as in it is possible that Johnwill arrive tomorrow).CB Segment Initial NO CB Total NumberMuseum 143 35 153 331Pharmacy 161 14 159 334Total 304 49 312 665But by far the most signiﬁcant improvement to thepercentage of utterances that satisfy Constraint 1comes by adopting a looser deﬁnition of ’real-izes’, i.e., by allowing a discourse entity to serveasCB of an utterance even if it’s only referred toindirectly in that utterance by means of a bridg-ing reference, as originally proposed by Sidner(1979) for her discourse focus. The following se-quence of utterances explains why this could leadto fewer violations of Constraint 1:(1)(u1) These “egg vases” are of exceptionalquality: (u2) basketwork bases supportegg-shapedbodies (u3) and bundles of strawform the handles, (u4) while small eggs restingin straw nests serve as the ﬁnial for each lid. (u5)Each vaseis decorated with inlaid decoration: In (1), u1 is followed by four utterances. Onlythe last of these directly refers to the set of eggvases introduced in u1, while they all contain im-plicit references to these objects. If we adopt thislooser notion of realization, the ﬁgures improvedramatically, even with the rather restricted set ofrelations on which our annotators agree. Now themajority of utterances satisfy Constraint 1:CB Segment Initial NO CB Total NumberMuseum 225 35 71 331Pharmacy 174 14 146 334Total 399 49 217 665And of course we get even better results by treat-ing sentences as utterances:CB Segment Initial NO CB Total NumberMuseum 171 17 39 227Pharmacy 168 7 36 211Total 339 24 75 438It is important, however, to notice that even un-der the best conﬁguration, at least 17% of utter-ances violate the constraint. The (possibly, obvi-ous) explanation is that although coherence is of-ten achieved by means of links between objects,this is not the only way to make texts coherent.So, in the museum domain, we ﬁnd utterancesthat do not refer to any of the previousCFsbe-cause they express generic statements about theclass of objects of which the object under discus-sion is an instance, or viceversa utterances thatmake a generic point that will then be illustratedby a speciﬁc object. In the following example,the second utterance gives some background con-cerning the decoration of a particular object.(2)(u1) On the drawer above the door, gilt-bronzemilitary trophies ﬂank a medallion portrait ofLouis XIV. (u2) In the Dutch Wars of 1672 -1678, France fought simultaneously against theDutch, Spanish, and Imperial armies, defeatingthem all. (u3) This cabinet celebrates the Treatyof Nijmegen, which concluded the war.Coherence can also be achieved by explicitcoherence relations, such as EXEMPLIFICA-TION in the following example:(3)(u1) Jewelry is often worn to signal membershipof a particular social group. (u2) The Beatlesbrooch shown previously is another case in point:Rule 1: if any NP is pronominalized, the CB isIn the previous section we saw that allowingbridging references to maintain theCB leads tofewer violations of Constraint 1. One shouldnot, however, immediately conclude that it wouldbe a good idea to replace the strict deﬁnitionof ’realizes’ with a looser one, because thereis, unfortunately, a side effect: adopting an in-direct notion of realizes leads to more viola-tions of Rule 1. Figures are as follows. Us-ing utterance=s, rank=gf, realizes=direct 22 pro-nouns violating Rule 1 (9 museum, 13 pharmacy)(13.4%), whereas with realizes=indirect we have38 violations (25, 13) (23%); if we choose utter-ance=ﬁnite, prev=suri, we have 23 violations ofrule 1 with realizes=direct (13 + 10) (14%), 32with realizes=indirect (21 + 11) (19.5%). Usingfunctional centering (Strube and Hahn, 1999) torank theCFs led to no improvements, because ofthe almost perfect correlation in our domain be-tween subjecthood and being discourse-old. Onereason for these problems is illustrated by (4).(4)(u1) A great reﬁnement among armorial signetswas to reproduce not only the coat-of-arms butthe correct tinctures; (u2) they were repeated incolour on the reverse side (u3) and the crystalwould then be set in the gold bezel.They in u2 refers back to the correct tinctures (or,possibly, the coat-of-arms), which however onlyoccurs in object position in a (non-ﬁnite) com-plement clause in (u1), and therefore has lowerranking than armorial signets, which is realizedin (u2) by the bridge the reverse side and there-fore becomes theCB having higher rank in (u1),but is not pronominalized.In the pharmaceutical leaﬂets we found a num-ber of violations of Rule 1 towards the end oftexts, when the product is referred to. A possi-ble explanation is that after the product has beenmentioned sentence after sentence in the text, bythe end of the text it is salient enough that thereis no need to put it again in the local focus bymentioning it explicitly. E.g., it in the followingexample refers to the cream, not mentioned in anyof the previous two utterances.(5)(u1) A child of 4 years needs about a third ofthe adult amount. (u2) A course of treatment fora child should not normally last more than ﬁvedays (u3) unless your doctor has told you to useitfor longer.5 DISCUSSIONOur main result is that there seems to be a trade-off between Constraint 1 and Rule 1. Allowingfor a deﬁnition of ’realizes’ that makes theCB be-have more like Sidner’s Discourse Focus (Sidner,1979) leads to a very signiﬁcant reduction in thenumber of violations of Constraint 1.3We alsonoted, however, that interpreting ‘realizes’ in thisway results in more violations of Rule 1. (Nodifferences were found when functional center-ing was used to rankCFs instead of grammati-3Footnote 2, page 3 of the intro to (Walker et al., 1998b)suggests a weaker interpretation for the Constraint: ‘there isno more than oneCB for utterance’. This weaker form ofthe Constraint does hold for most utterances, but it’s almostvacuous, especially for grammatical function ranking, giventhat utterances have at most one subject.cal function.) The problem raised by these re-sults is that whereas centering is intended as anaccount of both coherence and local salience, dif-ferent concepts may have to be used in Constraint1 and Rule 1, as in Sidner’s theory. E.g., we mighthave a ‘Center of Coherence’, analogous to Sid-ner’s discourse focus, and that can be realized in-directly; and a ‘Center of Salience’, similar to heractor focus, and that can only be realized directly.Constraint 1 would be about the Center of Coher-ence, whereas Rule 1 would be about the Centerof Salience. Indeed, many versions of centeringtheory have elevated theCP to the rank of a sec-ond center.4We also saw that texts can be coherent evenwhen Constraint 1 is violated, as coherence canbe ensured by other means (e.g., by rhetorical re-lations). This, again, suggests possible revisionsto Constraint 1, requiring every utterance eitherto have a center of coherence, or to be linked by arhetorical relation to the previous utterance.Finally, we saw that we get fewer violations ofConstraint 1 by adopting sentences as our notionof utterance; however, again, this results in moreviolations of Rule 1. If ﬁnite clauses are used asutterances, we found that certain types of ﬁniteclauses not previously discussed, including rela-tive clauses and matrix clauses with empty sub-jects, are best not treated as utterances. We didn’tﬁnd signiﬁcant differences between Kameyamaand Suri and McCoy’s deﬁnition of ‘previous ut-terance’. We believe however more work is stillneeded to identify a completely satisfactory wayof breaking up sentences in utterance units.ACKNOWLEDGMENTSWe wish to thank Kees van Deemter, Barbara diEugenio, Nikiforos Karamanis and Donia Scottfor comments and suggestions. Massimo Poesiois supported by an EPSRC Advanced Fellowship.Hua Cheng, Renate Henschel and Rodger Kib-ble were in part supported by the EPSRC projectGNOME, GR/L51126/01. Janet Hitzeman was inpart supported by the EPSRC project SOLE.4This separation among a ‘center of coherence’ and a‘center of salience’ is independently motivated by consid-erations about the division of labor between the text plannerand the sentence planner in a generation system; see, e.g.,(Kibble, 1999).ReferencesS.E. Brennan, M.W. Friedman, and C.J. Pollard. 1987.A centering approach to pronouns. In Proc. of the25th ACL, pages 155–162, June.J. Carletta. 1996. Assessing agreement on classiﬁca-tion tasks: the kappa statistic. Computational Lin-guistics, 22(2):249–254.H. H. Clark. 1977. Inferences in comprehension. InD. Laberge and S. J. Samuels, editors, Basic Pro-cess in Reading: Perception and Comprehension.Lawrence Erlbaum.S. Cote. 1998. Ranking forward-looking centers. InM. A. Walker, A. K. Joshi, and E. F. Prince, editors,Centering Theory in Discourse, chapter 4, pages55–70. Oxford.D. Cristea, N. Ide, D. Marcu, and V. Tablan. 2000.Discourse structure and co-reference: An empiricalstudy. In Proc. of COLING.B. Di Eugenio. 1998. Centering in italian. In M. A.Walker, A. K. Joshi, and E. F. Prince, editors, Cen-tering Theory in Discourse, chapter 7, pages 115–138. Oxford.B. J. Grosz, A. K. Joshi, and S. Weinstein. 1995.Centering: A framework for modeling the local co-herence of discourse. Computational Linguistics,21(2):202–225.J. K. Gundel, N. Hedberg, and R. Zacharski. 1993.Cognitive status and the form of referring expres-sions in discourse. Language, 69(2):274–307.I. Heim. 1982. The Semantics of Deﬁnite and In-deﬁnite Noun Phrases. Ph.D. thesis, University ofMassachusetts at Amherst.M. Kameyama. 1985. Zero Anaphora: The case ofJapanese. Ph.D. thesis, Stanford University.M. Kameyama. 1998. Intra-sentential centering: Acase study. In M. A. Walker, A. K. Joshi, andE. F. Prince, editors, Centering Theory in Dis-course, chapter 6, pages 89–112. Oxford.L. Karttunen. 1976. Discourse referents. In J. Mc-Cawley, editor, Syntax andSemantics7 - Notesfromthe Linguistic Underground. Academic Press.R. Kibble. 1999. Cb or not Cb? centering applied toNLG. In Proc. of the ACL Workshop on discourseand reference.D. Marcu. 1999. Instructions for manually annotat-ing the discourse structures of texts. Unpublishedmanuscript, USC/ISI, May.J. Oberlander, M. O’Donnell, A. Knott, and C. Mel-lish. 1998. Conversation in the museum: Exper-iments in dynamic hypermedia with the intelligentlabelling explorer. New Review of Hypermedia andMultimedia, 4:11–32.R. Passonneau and D. Litman. 1993. Feasibility ofautomated discourse segmentation. In Proceedingsof 31st Annual Meeting of the ACL.R. J. Passonneau. 1993. Getting and keeping the cen-ter of attention. In M. Bates and R. M. Weischedel,editors, Challenges in Natural Language Process-ing, chapter 7, pages 179–227. Cambridge.J. Pearson, R. Stevenson, and M. Poesio. 2000. Pro-noun resolution in complex sentences. In Proc. ofAMLAP, Leiden.M. Poesio and R. Vieira. 1998. A corpus-based inves-tigation of deﬁnite description use. ComputationalLinguistics, 24(2):183–216, June.M. Poesio, F. Bruneseaux, and L. Romary. 1999. TheMATE meta-scheme for coreference in dialogues inmultiple languages. In M. Walker, editor, Proc. ofthe ACL Workshop on Standards and Tools for Dis-course Tagging, pages 65–74.D. Scott, R. Power, and R. Evans. 1998. Generationas a solution to its own problem. In Proc. of the9th International Workshop on Natural LanguageGeneration, Niagara-on-the-Lake, CA.C. L. Sidner. 1979. Towards a computational theoryof deﬁnite anaphora comprehension in English dis-course. Ph.D. thesis, MIT.R. Stevenson, A. Knott, J. Oberlander, and S McDon-ald. 2000. Interpreting pronouns and connectives.Language and Cognitive Processes, 15.M. Strube and U. Hahn. 1999. Functional centering–grounding referential coherence in informationstructure. Computational Linguistics, 25(3):309–344.L. Z. Suri and K. F. McCoy. 1994. RAFT/RAPRand centering: A comparison and discussion ofproblems related to processing complex sentences.Computational Linguistics, 20(2):301–317.J. R. Tetreault. 1999. Analysis of syntax-based pro-noun resolution methods. In Proc. of the 37th ACL,pages 602–605, University of Marylan, June. ACL.M. A. Walker, A. K. Joshi, and E. F. Prince, editors.1998b. Centering Theory in Discourse. Oxford.M. A. Walker. 1989. Evaluating discourse process-ing algorithms. In Proc. ACL-89, pages 251–261,Vancouver, CA, June.B. L. Webber. 1978. A formal approach to discourseanaphora. Report 3761, BBN, Cambridge, MA. . pop- ular variants of centering theory, and using this corpus to automatically check two central claims of the theory, the claim that all utterances have a backward looking center ( CB) (Constraint. subsection of a leaﬂet was treated as a separate segment. We then identiﬁed by hand those violations of Constraint 1 that ap- peared to be motivated by too broad a segmenta- tion of the text. 2 Automatic. Amherst. M. Kameyama. 1985. Zero Anaphora: The case of Japanese. Ph.D. thesis, Stanford University. M. Kameyama. 1998. Intra-sentential centering: A case study. In M. A. Walker, A. K. Joshi, and E.
- Xem thêm -
Xem thêm: Báo cáo khoa học: "Specifying the Parameters of Centering Theory: a Corpus-Based Evaluation using Text from Application-Oriented Domains" pot, Báo cáo khoa học: "Specifying the Parameters of Centering Theory: a Corpus-Based Evaluation using Text from Application-Oriented Domains" pot, Báo cáo khoa học: "Specifying the Parameters of Centering Theory: a Corpus-Based Evaluation using Text from Application-Oriented Domains" pot