Centering in Japanese: A Step Towards Better Interpretation of 
Pronouns and Zero-Pronouns 
Shil,go ~l)lkada Norihisa Doi 
nti('higa,n¢}doi.cs.keio.ac.jp doi~doi.cs.kcio.ac.jp 
Department ol7 Computer Science, Keio University 
A|istract 
An extension of tile notion of "cen- 
tering" is described for illt erl)rcting zero- 
|)rOIlO\[lllS ~llld OVOFL l)l'O\[IOl\[llS in lia, Lll- 
r;dly oCcllrring Japanese l:exl,, hi l)rcvi - 
ous work, Oile ZeFO-\[)l'OllOIlll (~llcodo,q the 
t)ackward-looking center, with i)rollOllllS 
a, nd other zero-pronouns ha.ndh;d as if they 
were overtly expressed. All investigation 
is lnadc, and froln il, pronouli.q mid zoro- 
prollOtlliS al'(; couc, ludcd t,o t)(; liiorc snlionl, 
titan other ovorl, i/Otlll phl'a,~scs;, This eli. 
ablcs bei, ter intx;rprcl, ation of pronollllS aud 
ZOl'o-prOllOtlllS. 
i Introduction 
In ordc.r i,o twoid unnaturalness caused by redun- 
dant use of full noun phrases, l/rononlinal exprc'ssions 
arc used. In Japanese, therc ;ire basically two l,yl)cS 
ofl)ronolninal Cxl/ressions: the zero i)ronoul~ and the 
(overt) pr<mou,L Zero l)ronouns (mu I)c dcfiucd as 
fbllows \[Yoshimoto 8(3\]: 
A zero-pronoun is a noun phrase whi(:h is 
of an obligatory case ;rod which is not, ex- 
pressed but can Iw. understood I, hrough dis- 
COlll?SP. mid COlltexL 
There has been much work on han(lling zero- 
pl:O,,O,.l~, .~.'.h ,,,~ \[l<~,m,;y.,~l. S,5\], \[Vo~hi,,lot, o Sq, 
\[Walker 92\], mid \[N()moto !)3\]. Among; l.hcm, M. 
l(anmy~un~ showed in \[Kamcyama 85\] t.hat zero 
llrollOtlllS ill ,I;~I\])ltlICSC ,q(mi,ell('c,q could I)c illierl)rc.Lcd 
using a concept, called "(:Chic.ring" \[3oshi 81\]. In (,he 
centei'ing inodel, there is one cnt, il, y I,hal. all tltt, or- 
a, nce iliOS|, Colll, i'a\]ty concerns. This 0nl, it,y is rcfc:rro(I 
to as t, hc backw~ird-looking ceutor ((,'b), Any other 
(mtity appearing hi all Ill, tCl';tll('O is a. \[}.)r',var(l-lookillg 
center (Cf) which niny I)ccomc a (it) later Oil ill ttic 
discourse, Cfs arc ordered by grammatical flmctions 
according to the, Jr degrees of salience as follows: 
Topic > Subject; > Object/Object2 
> Others (Oblique, Possessor, etc) 
Kmneymn~ showed that t;he zeroq)ronoun corr(> 
Sllonds I,o the (3) in ;lalmnc.se. \[{ul., ill her ~tccotllll., 
it' there is more l,h&ll o11o zcrO-l)rOllOUn ill ;Ilk iiL- 
(,erance, only one of tlmm is t, lle Cb, and td\[ other 
Zel:O-l)rolloulls wcro handled just. as if they had been 
ov0rl;ly stal,e.d. |"lll'thCl'lllOl'(!, \])l'OllOUllS wcrc &L'-;o 
I.r(~nl.e(I as i\[' the cnl.ities had bc(m stated as 'ordi- 
llD.l'y' lIO/lll phl';tses. Ilut, t)V(!Ft pFOlIOllllS ~tFC used to,:) 
avoid tmllaturalncss, .just a.s Zel'O-l)l'OllOtlllS ~ll'C, ~111(\[ 
I, hcir ant, eccdeuts should be found. 
In this pa.pcr, ow:rl, pronouns, as well as zcro- 
l)ronouns, are interpreted l)y extending tlm notion 
of centering. B~lsically, entii, ies corrcsImnding to I, hc 
ZCl'O-pl?OIIO/lllS ~l.lld overt iil'OtlOtlltS ira2 all tre~d.cd as 
1oeing giwm lllOl'(~ attenl, ion than other entities in ~t 
S(~II{,(~IICC. ()lily ~h()sc \])l'()ll()llllS alld ZOl'O \])l'Oll()tlllS 
that arc of an interscntcntinl uaturc arc handled. 
So, those whose al~tcccdcnl, appears in I, he same scn- 
I,Cl/(:e as t,h(; l)rOllOltlillal ('lclllellt, i.e. inl, i'ascml,cl~tial 
anaphora, Hilt\[ I.hOSC W\]I()8(~ ant.ccc(lcnl; appears aft, or 
I, hc l)rOllOlllillai ClClllCllt,, i.e. (:al, nphora, arc Otll;sidc 
tim SCOl/C of this paper. 
In sccl, ion 2, the extcn(h'~d notion of c(mtcring the 
('cnt,cr l,is(. Model - is cxphdncd, lu section 3, a sys-- 
t.cm ilnl)lcmcnl, ing the (2cnt;er l,isi. Model is dcs(:ril)cd 
;m(l cvalua.l,cd. Conch.ling remarks are made ill see- 
lion 4. 
2 The Center List Model 
In this section, centering is cxtcndc'd to handle 
mult,il)Ic ZI~I!O-I)I'OIIOIIIlS ~ ~llld l, hcll \['url, her exl, cnd0d 
to handle over(, pronouns. Finally, the ordering; ot'm> 
t.ities for showing the degree of salience is &mcril)ed. 
115 1 
2.1 Zero-Pronouns 
In Kameyama's account, only one zero-pronoun 
encodes the Cb, and any otl~er zero-pronouns be- 
come Cfs, just as if they had been overtly ex- 
pressed in the sentence. In other words, when there 
are multiple zero-pronouns, only one of the zero- 
pronouns has any significance, and ally other zero- 
pronoun might as well have been overtly expressed. 
But, because entities become zero-pronouns in or- 
der to avoid unnaturalness due to redundancy, zero- 
pronouns can be said to be salient enough to be un- 
derstood without being overt. Iu effect, this means 
that a greater amount of attention is placed on them 
than entities that were overtly expressed. This is 
shown through an example. 
Taking her approach, some simple extensions are 
made to see how well the ordering of entities in cen- 
tering would work for multiple zero-pronouns. First, 
the antecedent for the Cb-eueoding zero-pronoun is 
chosen.as shown in \[Kameyama 85\]. Basically, this 
consists of choosing the entity with the highest de- 
gree of salience in the previous sentence. Then, the 
next most salient zero-pronoun according to the or- 
dering of degrees of salience given in the previous 
section is considered. The antecedent for this zero- 
pronoun is the most salient entity fi'om the previ- 
ous sentence which will not contradict any possible 
constraints. At this point, we only consider seman- 
tic constraints for excluding such sentences as "The 
desk ate fish" and contra-index constraints for ex- 
cluding such sentences as "Jack ate Jack." Any other 
zero-pronouns are handled in tile same manner. For 
example, the following discourse is examinedl: 
Example 1: 
(1) Taro wa Jiro to shokuji chuu de atta. 
Taro Top/Sub airo with meal during was 
Taro was h.avin 9 a meal wilh. airo. 
Cb: --, Cf: Taro > .}ire 
(2) @ Saburo we mikaketa. 
Sub Saburo Obj saw 
(Taw) saw SabTtro. 
Cb: Taro, Cf: Saburo 
(3) ~ • airo ni shoukaishita. 
Sub Obj Jiro Obj2 introduced 
(Taro) introduced (5'aburo) to Jiro. 
(4) ¢ q' Shokuji ni sasotta. 
Sub Obj meal Obj2 invited 
(Taw) invited (Sal, uro) to lhe meal. 
1 "¢5" denotes zero-pronouns, and Top, Sub, Obj, Obj2 de~ 
notes Topic, Subject, Object, and Object2, respectively. 
In sentence (1), tile Cfs are ordered as Taro > Jiro, 
since Topic is the most salient entity. In sentence (2), 
the entity with the highest degree of salience fi'o,n 
the previous sentence (Taro) is chosen as tile zero- 
pronoun's antecedent, and becomes the Cb, with 
Saburo becoming a Cf. In tile third sentence, after 
Taro is chosen as the subject of the sentence, since 
there is only Saburo left,, Saburo becomes the an- 
tecedent of tile object zero-pronoun, assuming that 
there is some sort of knowledge preventing Taro from 
becoming the object. 
After sentence (3), the ordering of noun phrases 
would be as follows: 
Taro (Cb) > Jiro (Cf- Obj2) = Saburo (Cf- Oh j) 
This means that sentence (4) is ambiguous, having 
tile following possible interpretations: 
(a) Taro invited Jiro to the meal. 
(b) Taro invited Saburo to the rneal. 
But, the preferred meaning is (b). So, this would 
mean that the ordering should be as follows: 
Taro > Eaburo > Jiro 
This example shows that when trying to interpret 
more than one zero-pronoun, the ordering of noun 
phrases according to Kameyama's acconnt may not 
be optimal. Of course, this can be rectified by chang- 
ing the ordering of the degree of salience so that Ob- 
ject is higher than Object2, and as noted later in the 
paper this will actually take place. But, suppose sen- 
tence (3) in Example 2 is replaced with the following 
sentence: 
(3') q5 '-b Jiro we shoukaishita. 
Sub Obj2 airo Obj introduced 
(Ta,,) i,,.t,~od,,ce~ ai,'o (to S,,b,,,'o). 
Even in this case, the interpretation of sentence 
(4) would not change 2. So, the ordering of zero- 
pronoun not being optimal, i.e. that zero-pronouns 
are more likely to become zero-pronouns again than 
overt noun phrases, would seeln to be the more log- 
ical choice. 
So, we propose that "entities that have become 
zero-pronouns are more centered in tile discourse 
than those that have been overtly expressed." There- 
fore, tile centering model has been extended to tile 
following two lists to handle entities (noun phrases) 
that appear in a sentence: 
~Althougb it should be noted that it doesn't seem to be ,as 
strongly preferred as before. 
1752 
(1) Center List . .. Entities in asentence that have 
becolne zero-\])rOllOtlns. 
(2) Possible Center List; •.. Entities in a sentence. 
that were overtly exl)ressed. 
2.2 Pronouns 
In Japanese, both overt and elided t)ronondnal 
forms exist. The elided l)rOl,Ominal forl'Ll (zero- 
pronoun) was discussed in the previous subsection. 
In this subsection, we will show how pronouns are 
handled within tile proposed nlodel. 
ill Kanleyatnlat~s acco/lnt, \])ronollltS do ltot elicode 
Cbs and can only become C\[~s. If overt pronouns 
are treated as entities that were overtly expressed 
(i.e. put in the Possibh'. Center I,ist), the \[blk)wing 
example will not l)e interpreted correctly a. 
l'~xanq)le 2: 
(1) Taro wa Jiro to hatnatshiteitat. 
Taro 'l'op/Sub Jiro with talking 
"l'(Iro was lalL"~ng wilh Jiro. 
CL:. -, 1)(211,:'l'atro > 3ire 
(2) (it \[lanatko we nfikaketa. 
Sub llatnako ()bj saw 
(Taro) saw llaua,(:o. 
eL: 'l'aro, PCI,: llanako 
(3) ~ 3ire ni kaatojo nituite hanashita. 
Sub 3iroObj2 her about talked 
(7'are) lalked to Jiro abottl hcr (Ilanako). 
CL: Taro, PCI,: Jiro > llauako 
(4) (l~ ¢l) Suki naatodcarl,. 
Sub Obj like is 
('/'.,',,) li~:~.,~ (Ha,,,~o). 
If this example is interl)reted with tit(', antecedent 
of kanojo (her) in sentence (3) in the Possible (.:eat- 
ter List, then tile interpretation would be 'Tatro likes 
3iro.' \[n order to obtain the preferred interpretation, 
the ordering of noun phrases should be as follows: 
Taro > llauako > .lifo 
'l'his example shows that pronoul,s are not n(~ccs- 
sadly at the same level as with other overt noun 
phrases. Ill other words, prollouns atre atl. at. level of 
attention higher than 'ordinary' noun phrases. This 
is especially true when considering tile fact that pro- 
llOllIIS are used to preven\[; nnl/aturathless dllc l.o re- 
dundancy, just as zero-prononiis are used. 
a'CL' strums for Center 1,ist., alld '|)('1,' stands for Possible 
Center Lisl,. 
So, we propose that pronouns be interl)reted at 
the same level as zero-pronouns as follows: 
C(mter List Model 
The entities in a sentence llelong to one of 
tile tbllowing two lists: 
(1) CeilteF List ... l';ntities that have. 
become zeroq)ronouns or overt l)rc ~ 
notlns. 
(2) Possible Center List ... Entities 
that were overtly expressed but ~tre 
not in the Center List. 
Entities ill the Center l,ist are more salient 
than those in the l)ossihle Center List, with 
tile exception of '\['opic. 
The exceptkm will be touched nllon in the next sub-. 
section. 
2.3 Ordering by Salience 
Next comes the problenl of ordering witl,in the 
Center l,ist and the Possible Center List. in other 
words, the difference in salience between pronouns 
(zero and ow;rt) and 'ordinary' noun phrases is 
shown by the Center List and the Possible Center 
last. l,'ntities in the Center List are lnore sMient 
than those in the Possible Center List. But, what 
about t, tl<" difference ill salie.nce within each list? 
In our model, the ordering is as follows: 
'l'opic > Subject > Object > Object2 > Others 
> Subject/Ol~ject/Ol~ject2 of subordinate clause 
> Others ill Sllbordill~Jt;(! clause 
The tirst line shows tile ordering of grammatical 
functions of the main verh. This line is basically the 
same as l(ameyama's ordering, except that Object is 
deemed to be more salient than Object2. This was 
because, after making some prelhninary evaluations 
of our model, Object was Rmnd to have at slightly 
higher degree of salience than Object2. 
The following two lines are for ally entities that 
atpl)car in subordinate (:l~mses. '\['here doesn't seem 
to be at clear cut difference between the Subject, Ob- 
ject, and Obje.ct2 of subordinatte clauses, so they are 
Ilcmdled at the same level. The difference between 
the lnain clause and any sul)ordinate clauses cap- 
tures the intuition that entities in the main chmse 
are. more s~dient than those in subordinate ones. 
There is one exception to the Center List Model. 
It is tile salience of the Topic in the Possible Center 
1153 
List. As can be surmised from the term itself, tim 
'Fopie is special in that the sentence contains infor- 
mation about the entity corresponding to the Topic. 
In other words, the sentence is usually about the 
3bpicalized entity. So, it was placed at. the same 
level as the Object in the Center List. 
3 Experiment and Discussion 
An experiment was done to show the effectiveness 
of the Center last Model in interpreting pronouns 
and zero-pronouns. A total of 160 sentences from 
the following four discourses were used: 
• "Ushikata To Yainanba" \[Tsul)ota 75\] (Japanese 
folklore - 70 sentences) 
• "Madogiwa No Totto-clmn" \[Kuroyanagi 81\] 
(Story--- 51 sentences) 
• "Yasei Doubutsu '.I'o Tomoni" (Newspaper col- 
Hllln) 
- "Baison" \[Obara 91\] (15 sentences) 
- "h'ie Wani"\[Obara 92\] (24 sentences) 
This section will first describe the simple imple- 
nlentation used in the experiment. Then, it is eval- 
uated (Table 1), followed by a eonll,a.rison (Table 2) 
with Kameyama's method. 
3.1 Implementation 
The implementation is kept simple to demonstrate 
the eft'eel of the Center List. Semantic constraints on 
the type of entity that a (zero) pronoun may refer to 
-- for example, the Subject of 'eat' must be animate 
-, and contra-index constraints for restricting combi- 
nations of eoreferring entities within a sentence - for 
example, tile Subject and Object of 'eat.' cannot be 
the same entity - are used. In addMon, a constraint 
concerning tire subject and identification 4 of Cbs in 
adjacent sentences is used \[Kameyama 86\], except it 
applies to each entity in the Center List of adjacent 
sentences as follows: 
Two zero-pronouns that appear ill the Cell- 
ter List of adjacent sentences should share 
one of the following properties (in descend- 
ing order of pret?rence): (l) id'entitication 
and subject, (2) identification only, (3) sub- 
ject only, (4) non-identification and not> 
subject. 
41(ameymna's terminology for l~mpathy \[Kuno 78\]. It 
shows the perspective from which an event is described. 
Of course, tire Center List and the Possible Cen- 
ter List by themselves will not be able to handle an- 
tecedents that arc not in the previous sentence. In 
order to solve this problem, an ad hoc approach was 
taken by adding the following two lists: 
• Past Center List ... Entities that have previ- 
ously been a zero-pronoun or an overt pronoun, 
but do not appear in the current sentence. 
• Noun List ... Entities that have never been a 
zero-pronoun or an overt pronoun. 
In order to avoid combinatorial explosion, the enti- 
ties that are held in these two lists are limited to 
those which appear in the previous three sentences. 
Each entity in the four lists is assigned a score to 
show its degree of salience. In other words, tbe score 
shows the possibility of beconfing a zero (or overt) 
pronoun in the next sentence. 
After morphological and syntactic analysis, the in- 
terpretation process is basically carried out as fol- 
lows: 
(1) Using the semantic constraints, possible an- 
tecedents for pronouns and zeroq)ronouns are 
found from the Center List, Possible Center 
List, Past Center last, and Nolln List. 
(2) Combinations of possible antecedents are made. 
(3) Contra-index constraints are applied. 
(4) Each combination is given a score as follows: 
(4.l) Compute the sum of the scores that each 
possible antecedent was given. 
(4.2) Give bonus scores according to the subject 
and identification constraint. 
(5) The combination with the highest score is ch{> 
sell as the combination with the inost probable 
antecedents. 
(6) The Center List, Possible Center List, etc. are 
updated. 
3.2 Evaluation 
Table 1 shows our results. Considering tile fact 
that t, he Center List Model itself handles only pro- 
notlns ~tnd gero-pronollns whose ~tnteeedents are 
found one sentence back, it shows promise since a 
very simple fi'amework is enough to achieve 76% ac- 
curacy. Also, though tile number of pronouns was 
small, tile percentage of correct interpretations was 
1154 
'l'al)le 1: llesulL of I,\]va, lu;ition 
All 
Anti 
Pro 
~ \['\['sulJota 75\] 
(2orr(~ct/Toi,al -~ 81/t05 
(7effect % 77% 
\[Kuroy:magi 81\] 
49/60 
\[()bar~ 91 \] 
l~/16 
8~% 
Correet/ToLal :/~l\[ 
(Jorreet (,~, 
75% 
r4/9t ,t215"e ~21~ 5 
81% 81% 80% 
Io,,,t..< 9 111 AI 
9/19 i5i~ 
.0/:}.2 =75% 1371170 \[ 81% 
7/s C<,rrod;/'i'ot~d # li 7/14 
I1 Correct ~ 
0/1 
50% 88% o% 
6/~i 0/1 0/0 
0/7t--1'1/:1-0 ~0% 47% 
~/~ ~/~ -I 100% 89% (Jorrect/'l'ot,;_ll # t CorreeL (~, o% 100% 
All 
Anl;1 
Ant1 -t- 
Pro 
All 7,el'o--pl'OilOIlllS ~ilid OVel'l, t)rOllOiiils 
i)ronolill.S ~illd zel'o-pr()ilOllllS whose ;inLecedelits ltl'e foulid Olil~ ,'-;t~ll\[,eiit'e t)il(:k 
Pl'OiiOlillS ~llld 7,ei'o.l)i'OliO/lliS whose ~mtecede.nts ~ll'(; fouiid ill(.)l'(; than ()ill? seliteliCe })aek 
All overL pl'OllOllllS 
jlISL under 90~. The i'(;lli~liilder of Ltlis slilisecl, ioii 
will ilia, ke some aii~dysis of I.he resuil,s. 
First, since l;h(; iilt, erl)relal,ioli o1' :t (zero) pro 
liOlill Iises the resuh; of I;he prc.vious ,¢;ell{ellC(?, <~rl'oi'- 
ehaillinl~" lllllSl, I;'(~ ciieck0d for. Error-.chililiilig oe- 
qliellL wi'oiig~ interlJreLai, iOli. Of 1.he :1.9 iil<;()i'l'eel, ili- 
1.0rpreta.l, ioiis, 11 (22<)/0) were due I,o I, his f~lc.l.or el 
I:her eon@etely (8) or parl, ially (:t). In l, lie case of 
\[Ob;ti'~t 92\], live oul; of t, he ten ~;rroi's were due t.o l.his. 
Along with error-chailiing, l, here is ~dso th<; pos- 
sihility of g(~tting the eorr(~el, iiil('.rl)r(~l,til, ion for the 
Wl'Oli~ l'eLISOll, i.e. ail el'rof in the lirevioils ,qelll,l)ll('e 
lii}ty c~tiise {LII iiil;el:i)l!eLitLiOll I,o lie correct;. ,qill(:e 
t, hore w(;re 49 ineoiu'i;t;L inl.erl)rel~ll.iolis , ill\[ 4{) il~lVO 
l, his poteni,i~d. \[lowew~r, Lliei'e wils olily oile citric of 
it fMse positive. 
The siniplieit, y of OlIF ilill)\]eirielii,al.ioli WilS also ii 
f~teLt)r in L|le Wl'Oli~ inl, t~.rpI:OLi/.i;iOllS. \'Vlio.il ii se.lll,f'llee 
iS ~l~ COiliplt'.X .qeill;CllC(;, Lhe sul)jeei: Dilly diili~r \[)etweeli 
dift~rent predicates. 
l:',x;tP.iple :l: 
<li 1 OliioI, Lll I,oorilii, (b7 T~iro we illit,sllkct,~l. 
S/It) think as Sill) Taro Obj fouiid. 
As q~i lho'ughC q~? \]'outed 7h~'o. 
<l>i trod <1>2 in l",xaNlple :l lil~ty or lll~ly ilol, \[)e the 
Slillle pc'rSOll. \[11 (till' sinil)le illtl)leitlellL~lLioli , illlil~,q,¢; 
C)lie el lihe coiisl;railiLS dCClii oLhei'wise, siich ca,ses :tt'c 
handled as I, ho s0+Dle. I/ut,, t.his led Lo 1,/ incorrecL 
iuterprcl, al,ions. 
A I~w iliOFO heuristic t'ill(~s, siich ~ls pl'(;fcrelices for 
l)~tr~tllel interl)rel,~ttiolis , would ~dso liltve r~tised the 
l)el'CC!lit{i.g~e o\[' eOl'l'ecL ii'il;erpreL;iLiOliS. 
As ettll \])e .q(?ell \['l'Olli Lhe low i)el;cellL~tge of COl'l'C('l; 
iilLerl)l'el, i~l,iOllS t'or (zero) pFOlIOUlIS> the higgest ci/llSO 
Of Wroli,g ini,ert)rel,al,ions is the hl.ek of a glohal dis- 
COllrSe liieohltili.<-;lti. This was the case for l:t wroligj 
iii{.erl)l.eLill, ioils. Fllri, hel'lllOl'C, of l,he 13, seven o('- 
(:ilrred wheii the discourse was inlx~rrupted hy a sin 
g;le SOlil,eilCl~ t, ll~lt, g~tve background iilforlil~ltion. 
Ai/oLlier cltuse for Wl'Ollg interpret,~Lions w~ts d/le 
Lo our tiled{:\[ I)eilig; I);/s(;d Oll scoring eomliina.tions of 
l)Ossibh; 0.1il, ecedelit, s. There is ~it;wlys i, he t)ossibil - 
ity of li/illtiple COllil)itia, Lic)lis hltvillg Lhe hest possible 
score, l",ighi> sileii c~ts,;s o(:ciirl'~;d ill Oilr ex~tl~iiii~l, ion. 
AiiiOlig,/,hose eighl; cils(!s, l, liere were llve cases where 
l.\]le (;oH:eel, ilii.erl)reL~t, ioii w~ls it\[iiOilg {lit! Lop COll-i- 
l)iiia~ioli,q. Alliong those five cases, tllerc were l, wo 
cases wliere the incorr~cl, iliterpr~;tatioli wi/s ehoseli. 
Finally, l,here wile only Olle ellse, where il zero- 
proliOllli did Dot h~tvo gre~li, er s~tlienc,; Lhltli ltil eni, iLy 
l, hil.t, at)l/elti'ed overtly. This occurred wheii ~tll elided 
()l)lique of ~t sul)ordinat, e clause was ordered as ha.v-. 
iilg }~;re~tLer s~/li(:n(:e l, hal~ ~l.li overL Oblique of ~t lil~tiii 
clause. 
3.:~ Comparison 
A comparison Lq lnzt(le ill Table 2 between our 
~q~l)l'o;~ch ((JenLer last) and Ka.ineyam~'s ~tppronch 
((.'enter), ,qinee \[(O.lllcy~tl\[l~'s al)pro~tch does not in- 
chide ow;rt proliolilis, t,\]it.~y were excluded froiil tile 
resulLs. Also, oilly the resulLs of ,qul)ject, object, 
~licl Ohjec, L:~ are IlSed. For ex;lliltll(; , l>he ()l)li<.lue 
l/,%q 
Cb 
Anti 
All 
Table 2: Comparisou 
II II Cb I Antl \[AU II 
Center List 92% \] 78% I 71% 
Center 88% 71% 65% 
Cb-encoding zero-pronouns 
Non-Cb-encoding zero-l)ronouns whose 
antecedents are one sentence back 
All non-eL-encoding zero-pronouns 
Object in passive sentences are excluded. 
There was one case where her approach was able 
to make a correct interpretation but ours could not. 
This, however, was a false positive. 
While all other differences between the two ap- 
proaches were cases where our approach was able to 
handle the interpretation but hers could not, three 
of the cases were not clue to a legitimate superior- 
ity of our approach. In one case, an error occurred 
due to error-chaining. In mini.her, the cause was 
the exclusion of the interpretation of the Oblique- 
Object in p~ssive sentences fi'om the evaluation of 
Kameyama's approach. The third case was the sin- 
gle false-positive tha.t occurred in the result of our 
approach, llowever, all other (sevell) cases were (\[tie 
to the salience ordering difference between the Ceu- 
ter List Model and Centering. 
Tim evaluation model was limited to a. very simple 
one so that the etfect of the Center List, i.e. the 
difference in ordering, would be apparent. Fronl tile 
comparison, the Center List Model can be said to 
order the possible antecedents more effectively than 
Kameyama's method. 
4 Conclusion 
In this paper, centering was extended to better 
interpret pronoulrs and zero-pronouns. It, extended 
the centering model to have two lists. The Center 
List holds entities that 'appeared' in the sentence as 
either an overt pronoun or a zero-pronoun. The Pos- 
sible Center List holds entities that overtly appeared 
in the sentence, excluding overt pronouns. 
A very simple implementation showed that 76% 
of pronouns and zero-pronouns could be interpreted. 
'\['he percentage goes up to 81% when considering 
only those whose antecedents are one sentence back. 
But, as the figures indicate, a more global frame- 
work, such as one descrihed in \[Grosz 86\], is needed. 
Acknowledgement 
The authors would like to thank CSK Corp. for 
providing the morphological and syntactic analyzer 
used in tile implelnentation. 
References 
\[Grosz 86\] Grosz, B. (1986). Attention, Intentions, 
and the Structure of Discourse. Computational 
Linguistics, Vol. 12, No. 3, pp. 175 - 204. 
\[Joshi 81\] Joshi, A. and Weinstein, S. (1981). Con- 
trol of Inference: Role of Some Aspects of Dis- 
course Structure Centering. Proceedings oflJ- 
CAI '81, pp. 385 387. 
\[Kameyama 85\] Kameyama, M. (1985). Zero Anapho- 
re: The Case of Japanese. PhD Dissertation, 
Stanford University l)epartnmnt of Linguistics. 
\[Kameyama 86\] Kameyama, M. (1986). A Property- 
Sharing Constraint in Centering. Pweeedings of ACL '86, 
pp. 20() - 206. 
\[Knno 78\] Kuno, S. (19781). Oanwa No nunpou, Tai- 
shuukan Shoten, in Japanese. 
\[Kuroyanagi 81\] Kuroyanagi, T (1981). Madogiwa No 
7'otto-chart, Koudansha, in J apanese. 
\[Nornoto 93\] Nomoto, T. and Nitta, Y. (1993). Re- 
solving Zero Anaphora in J N)anese. Proceedings 
of the 61h Conference of EACL, pp. 315 - 321. 
\[Obara 91\] Obara, 11. (1991). Yasei Doubutsu ToTo- 
moni. Mainichi Shinbun, Dec. 8, pp. 6, in Japa- 
nese. 
\[Obara 92\] Obara, II. (1992). Yasei Doubutsu To To- 
moni. Mainichi Shinbun, Jan. 19, pp. 6, in Japa- 
nese. 
\[Tsubota 75\] Tsubota, G. (1975). Oshikata ~lb Ya- 
manba. Nihon Mukashibanashi (1), Kaiseisha 
P, unko, pp. 93 98, in Japanese. 
\[Walker 92\] Walker, M., lida, M., and Cote S. (1992). 
Japanese Discourse and the Process of Center- 
ing. IRCS Report No. 92-14, The Institute for 
Research in Cognitive Science, University of Penn- 
sylvania. 
\[Yoshimoto 86\] Yoshimoto, K. (1986). Identifying Zero 
Pronouns in Japanese l)ialogue. Pweeedings of 
COLING '88, vol. 2, pp. 779 - 784. 
1156 
