<?xml version="1.0" standalone="yes"?>
<Paper uid="C00-1045">
  <Title>Pronominalization revisited*</Title>
  <Section position="3" start_page="0" end_page="307" type="metho">
    <SectionTitle>
2 Accounts of pronominalization
</SectionTitle>
    <Paragraph position="0"> In previous ace(rants pronominalization has been related to the idea of a local focus of ~d;tention: a set of discourse referents who/wlfich is in the center of attention of the speaker (e.g.</Paragraph>
    <Paragraph position="1"> Sidner (1979), givenness hierarchy (Gun(lel et al., 1993), centering theory (Grosz et al., 1995), RAFT/RAPId. (Suri, 1993)). Whereas (Gundel el; al., 1993) do not atteml)t to make their focus notion operationalizable, this has been attempted by fllrther develolmlents of centering.</Paragraph>
    <Paragraph position="2"> However these have mostly been applied to the pronoun resolution problem. In the following we discuss three versions of centering and show that their application to the pronoun generation problem is nevertheless linfited.</Paragraph>
    <Paragraph position="3"> Centering. Centering was developed to explain local discourse coherence; the extent to which it benefits pronoun generation is however not immediately clear, hi centering, 1We llSO the terms &amp;quot;discourse entity&amp;quot; and &amp;quot;refc'rent&amp;quot; synonymously in this paper.</Paragraph>
    <Paragraph position="4">  the discour.~e entiti(:s (:vok(:d in a (',(n'l;ain ui:terall(;e 'tt i ga'e c~dle(t fi)rward-looldng centers (Cfs). It is assumed i;lmt they are 1)ari;ially order(:d. As a major dei;erminant of the ordering, the gramma.tical fun(:t;ion hierarchy (roughly: SUI/.I~&gt;OIM&gt;OTIII~;IIS) has 1)een 1)r,')t)o,~ed. 13e-CallSe other fa.cl;ors afl'e(:ting th(: ord(:r have no(; 1)een (',lld)or;~ted in de(;ail, this ranking (as tit(' only Ol)erai, ionaliza/)le ha.n(lle) has 1)(:(x)m(: the si;:m(lard ra.nking in several comt)utational ai)1)\]ic~(,ions of (:entering. The 1)ackward-tooking center (hencefl)rth C1)) is a distinguished 1nerobet of (;lie CI:~, which is defined as the most highly ranked member of the Ct5 of th(' previous lltterallee 'u,i_ \] which is realized in v,i. The Cb is consid(:red as (;h(: \]o('al focus of ai;i;(:ntion. Centering sta,t('s two rules. Only the first; rule makes ~t (:brim at)()u(; l)ronominaliz~tion: ll! any (;lemen(; of the uti;eran(:e ui--1 is realiz(,,d in v,i as l)ronomh th(m l;he C\]) must l)e l)r()l~()minalized in 'it i as well. As lloted tw McCoy mid Struhe (\]999), this rule apl)lies only in (,h(; ease that (:we sul)seqllent lltterallCeS share Hl()r('~ (;}l~tll o11(~ ref('A'ellt~ m:td (;hal; l;h(: non-el) ref(.'r(,.nt is 1)ronominalized in (,he se(xmd ui:termme. \]}ut why (;\]fis non-el) referent is realized as a, l)ronomt is not given t)y (~h(' (:heory.</Paragraph>
    <Paragraph position="5"> Itowev(:r, f()lk)wing mot(: (;he sl&gt;irit of (xuli;ering tha.n the actual definition, ()he (:au understand (;he el) as (;he refin'ent which is preferC/d)lv l)ronominalize(t. General t)r(mominalizai;ion ()f the backward-looking center was in fact a claim of (mrly c(mtering, lmi; h~d 1;o l)e al)andone(1 because of (:olmter-evidenee from r(,'al discourse.</Paragraph>
    <Paragraph position="6"> Bnt the id(::r (,ha.t t)l'on()nfinaliztti;ion of the Ct) could 1)e a, m(:ans of establishing lo(:td discourse (x)herence is still 1)revalenl;. \]t has accordingly })een use(t l)y seine generation systems to (:ontrol 1)ronominalization e.g. in the IIA,;X sysl~em (O1)erlm~dcr et al., 1998), the el) is always realized as a pronomL Semantic centering. Centering is a.lso found in Dale (1990) as the method of t)ronominalization control. However, Dale's center detinii;ion differs from standard centering theory in i;hat it is defined semantically and not on the basis of a syni;aetie ranking. 2 This apl)roaeh has some appeal, espc.cially for generation, ix;cause it supl)ori, s the natural mo(hfla.rity bel;ween strate2In 1)ari;icular, \])al(: adol)l;s (;he l'estll(; o1' I;he aci;ion (Ienol:ed 1)3' ihe previous claus(; of a recil)(: as (;he center. ~i(: generation -wlfi(:h would determine t;he semantk: c(:nter for each uttera.llce and tactical g(:neration which dec.ides about granmmt:.ical fimetions.</Paragraph>
    <Paragraph position="7"> Functional centering. Finally, the cent;ering version suggest;ed \])y Stl'ul)e and Hahn (\]!)99) al&gt;l)ears to r(:veal an underlying discourse mechanism resl)onsil)le for centering: the information si;rH(:ture of an utterance (roughly the given-new i)ai;t(:rn) is (;he de('l)er reason for the ranking ()\[ the %rward-h)oking (:eni:(:rs. This l)ermii;s a generalization of sl;andar(l centering into a language-indel)ell(leld; l;heory eovering 1)oth trex: and fixed word-order languages. It is however then surprising that this result is not made maximld use of in the Sltb.sc,,(lttent generaISonorient, cA work of McCoy an(1 Strube (1999).</Paragraph>
    <Paragraph position="8"> Beyond centering. The questions wlfi(:h remain ol)en with all (;hree al)l)roat:hes - standard (;(;nterillg, S(:lll;~ll(;ic centering and fun(:-tional eentering- are: \].~.~ \~?\]ly are ill real texts a. lat%e nunl\])(:r of C1)'s not t)rononfinalized? _P21 \~q~y are non-C1)r(:ferents 1)ronominalized? or (:x1)ressc(l indel)(,.1Mently of centering: IP~I \~qty are in real texl;s a large nunfl)(,r of (tis(:our,s(: entities with an ani;ecedent in i;h(', previous utteran(:e not l)ronomina\]ized? FI?2~ \NqW can more tlmn one entity 1)e t)ronominalize(l in one ui;t(,rance? \].h'on~ a corlms-driven view, question \[~ is the larger prol)lem.</Paragraph>
    <Paragraph position="9"> ~/\[(;C()y all(t Stl'lt\])e (19,()9) were the first to sng;~est all al~ol;il~llln for ~ellel'~ttioll which solves t;h(,se problems. It was motivated by the ol)servation that ~t la.rge percentage of NPs which would have been realized by 1)ronouns using known algorithms, are in fact not, realized as pronomls in real text. They suggest that such NPs serve to mark ~time changes' in the discourse. Their algorithm aecordingly makes use of distance, context ambiguity mid telnl)ora\] discourse structure to decide about 1)ronolninalization. In our work, we have considered a corpus of a ditl'erent genre in which I, emt)oral cha.nge does nol, 1)bW a determining role: descript, ive (;exts. \Y=e 1)repose a new algorithm  that significantly simplifies the problem of pronoun choice. It is based on a new definition of the local focus, which views the discourse status of the antecedent as the major motivation behind focusing. The algorithm performs equally well when applied to McCoy and Strube's corpus of newspaper articles.</Paragraph>
  </Section>
  <Section position="4" start_page="307" end_page="308" type="metho">
    <SectionTitle>
3 Corpus analysis
</SectionTitle>
    <Paragraph position="0"> The algorithm we will present below has been developed in close relation to the MUSE corpus a corpus of museum exhibit labels a. The corpus is a collection of web pages of the Paul Getty Museum, pages from an exhibition catalogue, and pages froln a jewellery book. Typical characteristics are tile central role of inanimate referents in these texts, and the lack of temporal change thus providing an interesting countert)art to the newst)aper genre investigated by McCoy and Strube.</Paragraph>
    <Paragraph position="1"> With an overall set of around 5000 words, tile cortms contains 1450 NPs. Each NP has been annotated with respect to, among others, grammatical function, discourse status, gender, number, countability, and antecedent relationships. 23% of the NPs form reference chains (i.e. at least two mentions of one and the same referent in one text), the other 77% are only mentioned once. We have 101 different reference chains; the chain-fbrming NPs fall into 10\] discourse-new and 213 anaphoric NPs. In the following, we will only discuss the anaphoric NPs. 50% of the anaphoric NPs are realized as definite descriptions, 50% as pronouns. We distinguish between locally bound pronouns, which are determined syntactically (Binding Theory, (Choresky, 1981)), and which we expect the tactical generator to handle correctly, and pronouns which are not locally bound so-called discourse pronouns. We investigated possible col relations between the discourse 1)ronouns and semantic/pragmatic features of their context.</Paragraph>
    <Paragraph position="2"> The basic notions that we found were distance, discourse status of the antecedent, and grammatical function of the antecedent. All three notions need a precise definition.</Paragraph>
    <Paragraph position="3"> Distance. ~lb be able to determine the distance between a discourse entity and its antecedent, a precise determination of what counts 3UI{L: http ://www. hcrc. ed. ac. uk/'gnome/corpora as utterance unit is necessary. Following Kameyama (1998), we take as utterance unit the finite clause, l{elative clauses and con&gt; plenmnt clauses are not counted as utterances on their own. This means that we count clauses containing complement clauses or relative clauses as single utterances. 4,a The previous utterance is the preceding utterance at the same level of embedding.</Paragraph>
    <Paragraph position="4"> Note that we allow the treatment of clauses with VP coordination (subject ellipsis) as complex coordinated clauses as done in Kameyalna (1998), thus handling subject ellipsis as a discourse pronoun; our algorithm does not; insist on this view however.</Paragraph>
    <Paragraph position="5"> The following correlation between pronoun use and distance was tbund in our corpus: 97% of the pronouns have an antecedent in the same or the previous utterance.</Paragraph>
    <Paragraph position="6"> Discourse status. The information status of a discourse entity in an utterance is either given or new. We use these terms with an identical Lneaning as g~vn'nd and focus in Vallduvi (1993). Discourse status, as introduced by Prince (1992), is a similar but different notion: A discourse entity is discourse-old, if it has been mentioned before in the same discourse; it is discourse-new otherwise. All cases of givenness by indirect means like part-whole, set-member relationships, other bridging relations, inferences (Prince's inferrables, anchored and situationally evoked entities) are judged as discourse-new, thus taking into account only tile identity antecedent relationship. We share Prince's opinion that pronominalization has to do with discourse status, whereas definiteness has to do with information status.</Paragraph>
    <Paragraph position="7"> 66% of all short-distance discourse pronouns in the MUSE corpus refer to an antecedent which is in itself discourse-old.</Paragraph>
    <Paragraph position="8"> Subjecthood. The third strong correlation is the relation between pronoun use and the grmnmatical function of the antecedent. 63% of discourse pronouns have a subject as antecedent.</Paragraph>
    <Paragraph position="9"> The following table shows the overall distribution of antecedent properties for short-distance 4This deviates from Kameyama, who analyzes reported speech as separate utterance.</Paragraph>
    <Paragraph position="10"> 'SComplement and relative clauses consisting of inore than one tinite clause create their own internal level of focusing.</Paragraph>
    <Paragraph position="11">  discourse 1}ronouns and (shown ill 1}rackets) h),' short-distance definite descriptions.</Paragraph>
    <Paragraph position="12"> old new ,~,t)io{:{; a8% (22%) 25% (12%) ~ot s,,1,.i 28% (18%) .(}% (48%)</Paragraph>
  </Section>
  <Section position="5" start_page="308" end_page="310" type="metho">
    <SectionTitle>
4 Algorithm
</SectionTitle>
    <Paragraph position="0"> \] ~ased on these corpus s{,ndy resnl|;s, we {lefhle a new notion of the local focus -- the set of referents which arc awfilal)le for prononfilmlization.</Paragraph>
    <Paragraph position="1"> The local focus is ut}dated at each utterance boundary, and is defined as the set of referents of the 1)revious utterance which are:  (a) discourse-old, or (b) realized as subject.</Paragraph>
    <Paragraph position="2">  This set {:all theoretically {:ontain \]11{}re thnu one \].lt ill \]HOSt cas,, s, O0 a.(1 (\])) are o,,o and the same singlel;on sol;, which coul{l be seen as the well-known Cb. Thus sta\]ldar(l cent{~ring app{,,a\]'s as ~t spe{'ial case of {}m a\])l}roa{:h. This account means that newly introduced r(;t L {;l'C\]l|,S arc ll(}l; immediately l)ron{}mina\]ized i\]1 the following utterance, Ulfless they have bc{,.n introduced as subject--ml ol)servation made l}y Brenlmn (19!}8) and now confirlile(t with respect to our data also.</Paragraph>
    <Paragraph position="3"> The l}rOl}osed definition of the lo{:al focus ~,,ou-.. eralizes the t2){:using mechanisms assm~led i\]1 {x211i,el'\]llg all(t intro(hlces the discourse status ()f the antecedent as (}he lnain {:rii;eriol~ })ehi\]\]d th{~ l}\]onomi\]laliza.ti{}n decision. It is i\]\]t{;\]'{&gt;ting iT() lIOl,e I;ll:tl; \]X/i{:Coy all{1 Sl;rul}{; (1.9991 also nmke use of the discourse status of the ant(x:{,Aent; without mentioning it exl)lh:itly. For a. certah\] sul)sel; of' intrasentealtial anat)horic relal, ions ill amMguous cont(,xts they I)rOl}OSe \])ro\]\]on\]ina.1ization in case the antec(;dent would 1)c the 1)referred one in Stl'ut)e (1998)'s pr{}lloun resohllion algoritlnn. Because the set {}i' a.ntecedents is l'mlkcd there with resl)eCt to infbrm~tion status, this is identical with {mr proposal. Why tlmy do not use the discourse status as a general criBerion is not clear. We believe that the discourse status of the antecedent as pr{mominalization trigger is a general rule, in discourse Sell-l&amp;ii/;,ies.</Paragraph>
    <Paragraph position="4"> The central role of discourse sLat;us and subjecthoo{1 are in our opinion 1101; accidental. The Bwo nol;ions retlecl; tw{\] tyt)ical stra{,egies 1;o introduce a new referenl; inBo l;he (liscourse.</Paragraph>
    <Paragraph position="5"> Wc will assume here the mnm~rked inf(}rnmt;ion structure of an utterance: given - new. The subject usually is part of (or identical to) the 9i'uen. Let X i)e a certain referent; which is newly intl'oduced in utterance (ul), and referred to again in t;lle following ui~tera.nce 0121. In the first strategy, X is introduced in the new non-subject t)arl; of (u\].). And ill this l)ati;el'n the second lnention of X in (u2) is not pronominalized.</Paragraph>
    <Paragraph position="6"> In exalnple (1) given in Figure 1 tile local focus for ,,t ;el'a,,ce one el m0,1 : {t,.4; 'm..,in 'morns&amp;quot; is new in (\]1\].) and \]1ol; pronominalized in (112). The other typical strategy is where the referent is tirst mentioned ill a subject position. This is typical for a segment onset, or the beginning of ~ text,. Ofl;en this referent is given })y other lneans - - for example, l)y refhrence to a 1)icture., or to a r(;lated object. In example (2) of Figure \]., the second mention is i)rononfinal ized. ~\]'\]ms 1;11(} sul)jecI; position seems to timelion as creating a givemless allocation for the denotexl rcfercnl;. These two strategies roughly correspond with two types of thcnlatic developnlent identified in l)mm,q (197d).</Paragraph>
    <Paragraph position="7"> Parallelism. Our definition of l;he local tScus licenses 91% (62 of 68 pronouns) of all short-distance discourse pronouns in ore&amp;quot; corpus. Looking at tile pronomls violating the prol)osed accounl;, we nm.de ,,11 interesting observal;ion: n}osl; el l;heln occtlr ill conl;exts of strong t)arallelism. \'Vc call an anphoric NP ~/~,1~2 parallel if it has ml a.ntccedenl; 'l~q)l in the previous utter~Ill(;{;~ alld 'll,l) I alld '**,i12 \]rove Lhe 5alllC graummtical function. 1,k)l&amp;quot; work with real I;ex{;, il; is useful to inchlde cases whel'(~' 7L\]) 2 is a 1)osscssive or genitive NP inside a certain 'npa, and 'np\] and 'np:~ have the same gralnmatical flmction. Depending on the concrel;e function, we distinguish sub.ic{:t and object 1)arallelisln. Strong parallelism is a simulta.neous subject mid object para.lMism in two consecutive clauses. Strong i)arallelism always overrides the local focus criterion, mid allows tbr pronominalization of referents with discourse-new antecedents in nonsut)ject position. null The local focus definition refilled by the parallelism eff'ect is ml explanation for question P~ and a small portion of \[~\], but most cases of problem \[~ r(nnain open. q_'wo reasons for not prononfinalizing a reh~rent which is a nwanber of the local focus need to be considered: ~ alnbiguous context,  (1) (\[11.) Shortly after irzh, eriti~,.q the building in 1752, he commissioned th, e areh, iteet Pierre Conta, nt d'Ivry to renovate the main rooms.</Paragraph>
    <Paragraph position="8"> (u2) The engravings for these rooms , showing the wall lights in place, were reproduced in Dide~vt's Encyclopaedic, one of the principal works of the Age of E'nlightenment. (2) (ul) Scottish born, Canadian based jeweller, Alison Bailey-Smith, constructs elaborate and ceremoniaI jewellers} from irtd,ustrial wire.</Paragraph>
    <Paragraph position="9"> (u2) Her materials are often gathered from so'arces such as abandoned television .sets ... (3) ~i~ With attachments s'ach as an omtlav micvometer~ the microscope irtcovporates the latest seientitle technology of the mid-17OOs.</Paragraph>
    <Paragraph position="10"> (u2) The design of its era'ring gilt b~vnze stand was the heigh, t of the Rococo stifle ... (4) (u0) the table probably came from, the Trianon de Porcelaine , a small house built for th, e King's mistress, Madame de }l/\[ontespar~,, on the 9ro'unds of the Palace of Vet.sallies. (ul) This table's marquetry of ivory and horn, painted blue underneath, would have followed the house's blue-and-white color sehcrne, imitating blue-and-white Chinese porcelain, a fashionable aTI.d highly prized material.</Paragraph>
    <Paragraph position="11"> (u2) Blue-arM-white cevarnie tiles decorated the house, ...  discourse structure signalling.</Paragraph>
    <Paragraph position="12"> Ambiguity. Along with McCoy and Strube we argue that ambiguity with respect to gender/number influences the pronominalization decision: members of the local fbcus which have a competing referent (refbrent with similar gender/number) in some span to the left of the referent to be generated should not be realized as t)ronouns so as to minimize the inference load for the reader. However, not to allow pronominalization in all ambiguous context situations does not ~I)t)ear to be consistent with real texts (McCoy and Strube, 1999). In the MUSE cortms one third of all focal NPs occur in ambiguous contexts, one half of them is pronominalized, the other half is not. Two questions require a precise answer to use the ambiguity constraint in a generation algorithm: * Which set of 1)reviously mentioned referents or text st)an is taken into account tbr referents to be in competition? * Which referents are pronominalizcd despite an alnbiguous context? The answer is surprisingly simple: I/.eferents of the previous utterance which are not in the local fbcus do not disturb pronominalization, even if they have the stone gender/nmnber. Only if the actual referent has a competitor in the local fbcus, is pronominalization blocked. This is illustrated in Figure 1 with exmnples (3) and (4), respectively. In (3) the microscope is discourse-old and the only member in the local focus for (u2); the competing referents ocular micrvmctcr&amp;quot; mid technology are new and hence not local fbr utterance (u2). In (4), the local focus for (u2) is {the at, th, tl, e l,.o'asc} A slight improvement of the performance of the algorithm can be achieved by regarding the role of &amp;quot;heavy&amp;quot; nonrestrictive modification. hm\]uding the referents of discom'se-new NPs which are amplified by appositions or nonrestrictive relative clauses into the set; of' possible competitors improves accuracy slightly.</Paragraph>
    <Paragraph position="13"> Discourse structure signalling. It is now known that detinite descriptions (or more general overspecified NPs) signal the start of a new discourse segment (Passommau, 1996; Vonk et al., 1992). For most generation systems generate from an I/ST-like text; plan, discourse segments are naturally given. The only question fl'om the generation perspective is the degree of detail provided by the segmentation.</Paragraph>
    <Paragraph position="14"> Our algorithm gnome-up assumes that the discourse segmentation has already been specified. At each segment boundary, the local focus is set to nil, thereby disallowing pronominalization for all discourse entities of the first utterance in the segment onset.</Paragraph>
    <Paragraph position="15"> It is also well known that plmmed discourse with repeated phrases at the begimfing of a clause are seen as 'bad style'. Identical repeated pronouns at the clause onset are rarely found in expository and descriptive texts (2.6% of all discourse pronouns in our corpus). Hmnan writers usually avoid possibly dull lack of variation by employing various aggregation techniques.</Paragraph>
    <Paragraph position="16">  Let X 1)e a refl,'renl; I;o 1)e generated in Ill;l;erailCO (112), and focu,.s 1)e the' scl; of rc'h;reni;s of the 1)rcvious ul;l;eran(:(; (ul) which are  (a) discoursc-okl, or (b) realized as subject.</Paragraph>
    <Paragraph position="17"> (1) X has an antecedent beyond a segment boundary def description (2) X has an antecedent two or more ul;i;cranccs distant def (lescril)tion (:~) X hits ~Ul alll;(X'(xl(,ll(; ill (ll\]): ~lll(l (3a) X occurs in strong 1)aralM contc'xt 1)ronoun (31)) X C/ focu,.s (lcf dcscril)tion (3(:) X C .foc~z.s and * X has a coral)cling relbrcnt Y c focus dc'f description * X has a comp(~ting retba'cnt Y in (ul) amplitic(1 with appo- (lef dcscril)tion  sition or nonrestrictive relative clause * olso prollOllll The repeti|;ion 1)locking rul(; overri(lcs the 1)ronominalization suggesl;ed in (3c) to a definite description.  an aggregation trigger rather than ~ motivation for definite description generation. We hyl)ot\]&gt; (;size l;hat t;he at)l)ar(mt Kcquen(:y of (lelinite (lescriptions ill t)lmnm(l discourse has much to do with repetition blocking, but is used with respecl; to a very line-grained, 1)tel)ably genre-specific discourse, si;rtlCl;lll'e. Olle candidate for this is the, t(,mt)oral structure in newst)at)er ari;ieles proposed by McCoy and Sta'ul)e.</Paragraph>
    <Paragraph position="18"> When evahutting Ollr algorM m l, w(' only used tile pa.l'agr~I)h seglnenl;~ti;ion given in the corpus. \]{lit for g;etlel'al;ioll systellls, which usually sir(' not equil)l)Cd wit.h develol)ed a.ggrcgal.i(m mealtries, we have also made avai\]ablc a t)ronoun rci&gt; et,itioli blocking rule: If a discourse entity in the local focus has a nont)ossessiv(~ l)ronomilml an|;ecedelit, in'onolninalizal;ioli will 1)e J)loelC/cd at this l/line. Figure 2 SUll:lnlarizcs the algorithm.</Paragraph>
    <Paragraph position="19"> The presented pronominalization algorithm has been implelnented ill the reusable module gnome-np, gnome-np consists of a colnponent for discom'se model lnanagement and one for NP form determination, it is designed to 1)e plugged ill ~~:\['1;(;1' text 1)lanning, coneeI)tualizati(m, and sentence plalming, trot 1)etbre tactical generation.</Paragraph>
  </Section>
class="xml-element"></Paper>