Corpus-based Development and Evaluation of a System for Processing 
Definite Descriptions 
Renata Vieira 
Universidade do Vale do Rio dos Sines 
Av. Unisinos 950 - Cx. Postal 275 
93022-000 S~o Leopoldo RS Brazil. 
renata@exatas, unisinos, br 
Massimo Poesio 
University of Edinburgh 
HCRC and Informatics 
Edinburgh, Scotland 
Massimo. Poesio@ed. ac.uk 
Abstract 
we present an implelnented system for processing 
definite descriptions. The system is based on the re- 
sults of a corpus analysis previously reported, which 
showed how common discourse-new descriptions 
are in newspaper corpora, and identified several 
problems to be dealt with when developing compu- 
tational methods for interpreting bridging descrip- 
tions. The annotated corpus produced in this ear- 
licr work was used to extensively evaluate the pro- 
posed techniques for matchiug delinite descriptions 
with their antecedents, discourse segmentation, rec- 
ognizing discourse-new descriptions, and suggest- 
ing anchors for bridging descriptions. 
1 Motivation 
In previous work (Poesio and Vieira, 1998) we re- 
ported the results of corpus annotation experiments 
in which the subjects were asked to classify lhe uses 
of delinite descriptions in Wall Stree Journal arti- 
cles according to a scheme derived from work by 
Hawkins (1978) and Prince (1981 ) and including 
three classes: I)II{I'~CT ANAPltORA, I)ISCOURSE- 
NEW, and I~RIDGING DESCI{II'TION (Clark, 1977). 
This study showed that about half of the time, deli- 
nite descriptions are used to introduce a new entity 
in the discourse, rather than to refer to an object al- 
ready mentioned. We also observed that our sub- 
jects didn't always agree on the classification of a 
given delinite; the problem was especially acute for 
bridging descriptions. 
In this paper, we present an implemented system 
for processing delinite descriptions based on the re- 
suits of that earlier study. In our system, techniques 
for recognising discourse-new descriptions play a 
role as ilnportant as techniques for identifying the 
antecedent of anaphoric ones. The system also in- 
corporates robust techniques for processing bridg- 
ing descriptions. 
A fundamental characteristic of our system is that 
it was developed so that its perfomlance could be 
evaluated using the annotated corpus. In the papm; 
we discuss how we arrived at the optimal version of 
the system by measuring the performance of each 
method in this way. Because of the problems ob- 
served in our previous study concerning agreement 
between annotators, we evaluated the system both 
by measnring precision/recall against a 'gold stan- 
dard' and by meastu'ing the agreement between the 
annotation it produces and the mmotators. 
2 General Overview 
At the moment, the only systems engaged in se- 
mantic interpretation whose performance can be 
evaluated on l'aMy unrestricted text such as the 
Wall Street Journal articles are based on a shallow- 
processing approach, i.e., that do not rely on exten- 
sive amounts of hand-coded commonsense knowl- 
edge (Carter, 1987; Appelt, 1995; Humphreys et al., 
1998). I Our system is of this type: it only relies 
on structural information, on the infommtion pro- 
vided by pre-existing lexical sources such as Word- 
Net (Fellbaum, 1998), on minimal amouuts of gen- 
eral hand-coded iuformation, and on information 
that can be acquired automatically from a corpus. 
Although we believe that quantitative ewduations of 
the performance of a systeln on a large number of 
examples are the only true assessment of its perfor- 
mance, and therefore a shallow processing approach 
is virtually unavoidable for implemented systems 
until better sources of commonsense knowledge be- 
come available, we do know that this approach liln- 
its the performance of a system on those instances 
of definite descriptions which do require common- 
sense knowledge for their resolution. (We grouped 
these in what we call the 'bridging' chtss.) We 
I Most systems participating in the Message Understanding 
Conference (MUC) evaluations are customized to specific do- 
nmins by adding hand-coded commonsense knowledge. 
899 
nevertheless developed heuristic techniques for pro- 
cessing these types of delinilcs its well, which n/ay 
provide a baseline against which the hains in perfor-- 
nlanco (\]lit to tile llSO oi: COlllnlOllSellse knowlodoe 
can be assessed more clearly. 
Our system attempts to classify each deli- 
nitc description as either I)IRIX:T ANAI'ItORA, 
I)IS('.()UI~,SIT,-NI';W, all(t IgRII)GING I)lv.S(21{II'TION. 
The lirsl chlss includes deihfite descriptions whose 
head is identical to thai o1' their antecedent, as in 
a Iiouse ... lhe house° The second includes del L 
inile descriptions that refer to objects not alma@ 
mentioned in the text and ,lot rclated to any such 
object. (Some of these definite descriptions refer 
to objects whose existence is widely known, such 
as discourse-initial references to lhe i)ot;e; other in- 
stances of discourse-new descriptions refer to o1:> 
jects thai can be assumed to bc unique, even if unfa~ 
miliar, such as lhe.filwl woman lo climb all Scollish 
Mum'os.) lqnally, we classify as bridging descrip- 
tions all dclinitc descriptions whose resolution de-. 
pen(Is on knowledge of relalions between objects, 
such as delinite descriptions thai refer to an object 
rehlted 1o an entity ah'eady introduced in the dis-- 
com'se by a relation other than identity (Prince's 
'inlerrables'), as in the flat.., the living, room; and 
de/tulle descriptions that refer an object aheady m 
troduced, but using a different predicate, as in Ihe 
car...lhe vehicle, hi addition to this chlssitica- 
lion, the system iries to identify the antecedents 
of anaphoric descriptions and the anchors (Fraurud, 
1990) of bridging ones. Accordingly, we developed 
three types of heuristics: 
, for resolving directly anaphoric descriptions. 
These iuclude heuristics for dealing with see- 
mentation and to handle modificatiou. 
• for identifying discourse-new descriptions. 
Some of these heuristics attempt to recognize 
semantically lunctional definite descriptions 
(Hawkins, 1978; Loebner, 1987), whereas oth- 
ers try to recognize definite descriptions that 
are anchored via their modification (Clark and 
Marshall, 1981; Prince, 1981). 
,, for identifying the anchor of a bridging de- 
scription and the semantic relation between the 
bridging description and its anchor. WordNet 
is accessed, and heuristics for named entity 
recognition were also developed. 
The final configuration of the system was arrived 
at on the basis of an extensive ewfluation of the 
hemislics using Ihe corpus annotated in our previ- 
ous work (Poesio and Vieira, 1998). The evaluation 
was used both io detemfine whMl version o1' each 
heuristic worked belier, and to identify the best o1", 
der in which to try them. 
The corl)us we used consists of 3d. texts frolu/he 
Peun Treebank I included in the ACIJl)CI (\]\])-rolu. 
20 of lhese texts were treated its 'training,, corptls'; 
this corpus contains 1040 (lclinite descril)tions, ()l' 
which 312 arc anaphoric, /192 discernso-.new, and 
204 bridging, id more texts were used as 'test 
corpus'; fllesc include d64 delinile description,% of 
which 154 haw: been classified its anaphoric, 218 as 
discourse-new, and 81 as bridging. 
3 Tlhe Hem~istics And Their Perfbrmance 
3.1t /Resolving Anaplhorie Detinites 
We discuss heuristics for two sut)problcms of lhe 
lask o\[ resolving anaphoric dclinites: limitin?; ihe 
accessibility of discourse entities (seomcnlation), 
and |aking into accotln/ 1t4e information given by 
pre- and post-modiliers. See (Vieira, 1998) for a 
discussion of tile other heuristics used by the sy> 
tonl. 
Segmentalion In 
l i IE-spans limited 
MI",NTS t1131 lllay 
general, discourse entities have 
to pra?matical ly delermined ,Slit;.. 
be nested (see, e.g., (Rcichman, 
1985; Grosz and Sidner, 1986; Fox, 1987)). E.g., 
in our corpus we found that about 10% of direct 
anal~horic detinite descriptions have more than one 
possible antecedent if segmentalion is nol taken into 
account (Vieira and Poesio, 1999). Recognizing 
the hierarchical structure of segments in a text is, 
howevm; still pretty umch an open problem, kS it 
involves reasoning about intentions; 2 better results 
have been achieved on the simpler task of 'chtlnk- 
ing' the text into approximate segments, generally 
by means of lexical density measures (Hearst, 1997) 
In fact, the lnethods Io limit the lifespan of dis- 
course entity we considered for our system were 
even simplel: One type of heuristics we looked 
at are window-based techniques, i.e., considering 
as potential antecedents only the discourse entities 
within fixed-size windows of previous sentences, al- 
lowing however for some discourse entities to take 
a longer life span: we call this method LOOSE SEG- 
MENTATION. More specifically, a discourse entity 
is considered as potential antecedent for a definite 
2See, howeve,, (Marcu, 1999). 
900 
doscril)iiol~ w\]l(:li lhc mi{oc~:dclii'~; head i~ idcniic~ll 
lo lhc doscril)lioli~s licad, nnd 
o lho poicnlial anlcc:cdoniL <', (li,,4ailcc lroul llic do-. 
,smrip/ioii b; \vidiin Ihc es/ab\]ishod window, or 
c;i sc 
o ihe ix)lcnliai aulcccdcnl is il~;cll-~t stil)~;t:ClllCli{ 
illCilliOll, or cls(: 
o ill~; (lciilfiic dcscrif)iitni alid lhc 4iiicc:cxJoiii ;llC 
ktcnllcal Ni's (iiiohidhig ilic ariiclc). 
Wo also coiisido'rcd mi c'vcii shill)lor I',i,;(:i,.NCY 
hc:uristic: fllis involw:s kt:cl)iu ?, a l~ibl,: iiidc:xcd l)y 
ihc" hc, ad.'; o\[ i)()ic-ntial nuli'x:cd(;lli,<;, ,';llch dial the cn- 
irj for iiOllil \[\I c(;iliahl<~ ihc iilc!cx oi ilic I:isi t)cciil (. 
rciio(', t)\[:ilii 4ui(tcc'.dciii wilh licad N. i:iiulliy~ wc (',oii- 
,<;i<icri'd moiul)hiaii{)ny, <fl~;('l,, c. lllllitiii ;iiid rc~(:cuc3,. 
Tli(~ be',<;1 rl'!;ull,<; wi;rc obi;ihicd wilh a couit)i- 
iia{i()n ()1 ilic: rccmicy mid ,<;c;?,iii('.ul~liion hcuri<~lic,<,: 
.iu,'a o,c polciiii41 anicccdclli \[or CilCii dJl\]orc, iH lioad 
liOtlil iF, awlilal:)lc Jor r~soluiion> llw la!d oC('llrfOllCO 
()t ihai hc41d ilOlili. Tilt; rc,solutic)u ,~lill re.six:oh; lhc 
sc:Oill{:ill;lliOii h?llrislic (I()()SO vcrxioli). 'I'll(; rc:(:;ll\] 
(J:,), i)ro(;i,';ioit (P) all(l IL-illt:ilStlrc (l;) ft:stlllt; Io\] l\[lo 
lwo hcurislics ;irc i)rc:sc;ntt:d hi 'l\]iblc I. 
( kmfl>iucd lic/u i:-ilics I l> I i' I I: 
"I £CIIICIICCS -I rt}ct~llcx/ 1 "/~7-()()(A' KI.77<A ' S i. '!'7!(/< 
{~ SCIIICIICCS I rt'CCilCy I "li.~I4'J S,I\]I)()<,/~ ~gi.'~7'/<, 
'lhl)lc: I: (:OlilI)iilii/~t Ioo~;r ,'-;t;l'JllC:ili~lliOl! ;llltl i'(; 
CCiiCy lict;ri',dics 
The vcr<<doli with liighc:r I,' value: iu 'lhhlo I (4- 
.~;ciitt;liOC window i)his rt:ct:llcy) \vat.; dlO.'4t;ii aud 
u,.4od hi ill(" icsl,<; di,~mus.sod hi/he icst <ff lliis ,<;cc/iou. 
/Vom~ Mod~fie#;~' lll 9,onoial, wlioii niaichhi~,~ a (Ioli- 
liito doscriplion wilh a i)olt;utial aillocodOlit tho in- 
\[orillatiOli provided hy lho pfCllOlllhlaI ;Jild lhc l)OSl - 
noinifial pni( of lJic ll<)tili i)hraso also ll;i<<; lo I)o lakon 
iltto aoco/inl: SO, lor O×alnplo, ~t i)\[ue cdv Cailll()t 
SClVO ~lS the anlcccdonl for l/re red cas, of lhe hottxe 
on the 1(',./7 for fl#e house on lhe righ#. 'ltiking l)ropor 
care of the scniantic contril)uiion of these prcnlod- 
iliers would, in ~oiloral, roquiro OOliimOllSt'nso roa- 
lhc <'-;laiidard dchnihOliS of precision aild recall ffoill ilif()i- 
nlalion rt:lricvai \VOfC used: R :: illlnlbc, r of ottIocIS ('f type A 
corrcclly idcnlilicd I)y the system / tolal lltlmbcr of objccls o1" 
lyi)e A, I ) = llunfl~c:r (fl" corrccl ideniilicali(ms of oh.iccls of lypc 
A / l\[)I~11 nund}cf of objects of lyl)c A identilicd by file system, 
1;:~ RI)/1~,+1 ). 
,<;olihiF, ~ Jk)l- ilic; lilonit:lil, WC unly dcvclolxxl hc.uiiv 
iic sohilioii.~> to ihc prol)lcni> hichidhil,.: 
allowiug all anlcccdcnl to n-ialch with a dcli 
nik; dtJscripl\]oli iI lho l)lcn\]odificrs ol: tile d('- 
scrip/ion illt: ~1 ,'4ttl),'4t;l t)l lhc l)rc',niodiiiurs (>1 
ihu, aiilcccdciii. 'rhi,<~ heuristic deals wiih &'l t- 
hiiiot-; whicli mmlain io~,s inlornmlion lhan lilt; 
;tillt;t;c;dc;li/~ Silch ~i,<; fill 0/(\[ VicloFidn hotlmd... 
/hu /IOlt,~d~ ~lli(t t)rt:vt;il/S Illli/c;ho,~; ,~;uch zis ih¢ 
/)!/A'iH(tXX COIIHH\[Illi\[Y... lhc fOlllL~(t#; IIIf)F(f (l(;-. 
iivixl I)l(/(:k l)oliii(:al (:ommus~i/y. 
~, allowing a n(m-i)rolnodilicd auteccdoilt lo 
lllaich wilh lilly SalllO head definite. Tllis t;cc 
oud ht;ulixlic (!c'~il~; with dc, iinitc'.~ lhai (:onlilhl 
nddilioii;il iil\[ornlaiion, such ;l!', (~ check°., llm 
Io:>7 check. 
The ll'MIIi,~; o\[ ~)lu i)roniodiJit;r in,lithium, ~ll~to-. 
rilhni art: l);cs(:iilod hi Table 2. hi lhai Tal)lc; we: al,~;o 
.<;how lhc: fO.~lilb; oblahlod wilh a niodiliod lnatchhig 
algorillnn hicludhil,> ~i ihird rulo> thai allows ;i lm; • 
modified nnlLc;Ccdoni io nialch with a dclhfiio wl~o,~;o 
xcl ol i)rc-liiodilic;rs i,s a SUl)Oisol of lhc s('i of iuodi- 
lit'r.~ o1: lho aiilccc, donl (Jill claboralion (ff rule 2). \Vc 
Ic~dcd ciicli of lllcso lhrt;o hourislics alollt: mid ihcir 
coiiil)iiialit)l/,<;. (Tlic. Iourih liliC .~hnl)ly lc;i)~',;il,~ lli~: 
lc, sldis showu in 'lhl4c 1.) 
Alllcccdci/\[s sclcclion 14 1 1' 
I. Aiil -,';till )csc'- subscl ()9.197</< , 91.21% 
2. Ani-cinply 3_'i. 12% <$8.2()</~ 
3. Ant suhsct/I)csc, sci ().I.'t4{A 
i,i,a 7 0.i~<: ,,.> 7s-'~U 15/.H% 
1 and 3 "15.96% ~1.13% 
None 7g.. ~2% g .03% 
Tal)h" ?,: l:,valualioll of ihc houristics 
cqlIoII (Vt'l'Sit)ll 1 ) 
I I: 
7g. I "%4 
I 0',I.~59; 
'iq.,'11(/< ~J..,!:,1% I 
,*~ I. I (t/< 
~(. o% 
Ik)r laCJm)dili- 
The I)csl inccisiol\] is achieved by lhc niatchin~,, al. 
gorillnn Ihm does not allow for new information in 
tilt; anaphoric expression, but ilic best results ovc> 
all arc again obtained by combining rule I and rule 
2, alfllough oilhcr 2 or 3 works equally wcll when 
combinod with 1. 
Ovc#wll results .for amqUioric Ue.finite Uescriplions 
'lk) sumnm|'izc, lhc version of the system that 
achiow, s the best rcsulls as far as anal)hOl'iC dclinilc 
descriptions arc concerned includes : 
I. combined scgmmHation and roconcy~ 
901 
2. 4-sentence window, 
3. considering indefinites, definites and posses- 
sives as potential antecedents (Vieira, 1998), 
4. the pmmodification of the description must be 
contained in the premodificatiou of the an- 
tecedent when the antecedeut has no premodi- 
tiers. 
3.2 Heuristics for Recognizing Discourse-New 
Descriptions 
As mentioned above, a central characteristic of our 
system is that it also includes heuristics for recog- 
nising discourse-new descriptions (i.e., definite de- 
scriptions that introduce new discourse entities) on 
the basis of syntactic and lexical features of the 
noun phrase. Our heuristics are based on the dis- 
cussion by Hawkins (1978), who identified a num- 
ber of correlations between certain types of syntac- 
tic structure and discourse-new descriptions, par- 
ticularly those that he called 'unfamiliar' definites 
(i.e., those whose existence cannot be expected to 
be known ou the basis of generally shared knowl- 
edge), including: 
• the presence of 'special predicates':4 
- the occurrence of pre-modifiers such as 
first or best when accompanied with full 
relatives, e.g., the .\[irs't peJwon to sail 
to America (Hawkius calls these 'un- 
explanatory modifiers'; Loebner (1987) 
showed how these predicates may license 
the use of definite descriptions in an ac- 
count of definite descriptions based on 
functionality); 
- a head noun taking a complement such 
as the fact that there is li\['e on Earth 
(Hawkins calls this subclass 'Nt' comple- 
ments'); 
• the presence of restrictive modification, as in 
the inequities of the current land-ownership 
system. 
Our system attempts to recognize these syntactic 
patterus; in addition, it considers as unfamiliar some 
definites occurring in 
4This list was developed by hand; more recently, Bean and 
Riloff (1999) proposed methods for autolnatically extracting 
fl'om a corpus such special predicates, i.e., heads that correlate 
well with discourse novelty. 
• appositive coustructions (e.g., Glenn Cox, the 
president of Phillips Petroleum Co.); 
* copular constructions (e.g.,the man most likely 
to gain custody of all this is a career politician 
named David Dinkins). 
In our corpus study (Poesio and Vieira, 1998) 
we found that our subjects did better at ideutify- 
ing discourse-new descriptions all together (K=.68) 
than they did at distinguish 'unfamiliar' from 'larger 
situation' (Hawkins, 1978) cases (K = .63). This 
finding was confirmed by our implementation: al- 
though each of the heuristics is designed, in princi- 
ple, to identify only one of the uses (larger situation 
or unfamiliar), they work better when used all to- 
gether to the class of discourse new descriptions. 
The overall recall and precision results for the 
heuristics for identifying discourse new descriptions 
are shown in Table 3. In this Table we do not distin- 
guish between the two types of discourse-new de- 
scriptions, 'unfamiliar' and 'larger-situation'. The 
column headed by (#) represents the number of 
cases of descriptions classified as discourse new in 
the standard annotation; + indicates the total num- 
ber of discourse-new descriptions correctly identi- 
fied; - the nmnber of errors. These results are for the 
version of the system (version 1) that uses the best 
version of the heuristics for dealing with anaphoric 
descriptions discussed above, and that doesn't at- 
tempt to resolve bridging descriptions. 
I Discourse new l# \] + l- I P" I P I F I 
Training data 492 368 60 75% 86% 80% 
Test data 218 151 58 69% 72% 70% 
Table 3: Evaluation of the heuristics for identifying 
discourse new descriptions 
3.3 Bridging Descriptions 
Bridging descriptions are tile class of definite de- 
scriptions which a shallow processing system is 
least equipped to handle, and therefore the most cru- 
cial indicator of where commonsense knowledge is 
actually needed. We knew from the start that in 
general, a system can only resolve certain types of 
bridging descriptions when supplied with an ade- 
quate kuowledge base; in fact, the typical way of 
implementing a system for resolving bridging ref- 
erences has been to restrict tim domain and feed 
the system with hand-coded world knowledge (see, 
e.g., (Siduei; 1979) and especially (Carter, 1987)). 
902 
Furthermore, the relation between bridging descrip- 
tions and their anchors may be arbitrarily complex 
(Clark, 1977; Sidnm; 1979; Prince, 1981; Strand, 
1996) and our own results indicate that the stone de- 
scription may relate to different anchors in a text, 
which makes it difficult to decide what the intended 
anchor and the intended link are (Poesio and Vieira, 
1998). Nevertheless, we feel that trying to process 
these definite descriptions is the only way to dis- 
cover which types of commonsense knowledge are 
actually needed.. 
We began by developing a classilication of bridg- 
ing descriptions according to the kind of informa- 
tion needed to resolve them, rather than on the ba- 
sis of the possible relations between descriptions 
and their anchors as usually done in the literature 
(Vieira, 1998). This allowed us to get an idea 
of what types of bridging descriptions our system 
might be able to resolve. We classified definite de- 
scriptions as follows: 
• cases based on well-delined lexical relations, 
such as synonymy, hypernymy and meronymy, 
that can be found in a lexical database such as 
WordNet (Fellbaum, 1998)-as in theflat.., lhe 
living room; 
• bridging descriptions in which the antecedent 
is a proper name and the description a common 
noun, whose resolution requires some way o1' 
recognizing the type of object denoted by the 
proper name (as in Bach ... the composer); 
• cases in which the anchor is not the head noun 
but a noun modifying an antecedent, as in the 
compaw has been selling discount packages 
... the discounts 
• cases in which the antecedent (anchor) is not 
introduced by an NP but by a vl', as in 
Kadane oil is currently drilling two oil wells. 
The activity... 
• descriptions whose the antecedent is not ex- 
plicitly mentioned in the text, but is implicitly 
available because it is a discourse topic-e.g., 
the industo, in a text referring to oil coml)a- 
nies; 
• cases in which the relation with the anchor is 
based on more general commonsense knowl- 
edge, e.g., about cause-consequence relations. 
We developed heuristics for handling the first 
three of these classes: lexical bridges, bridges based 
on names, and bridges to entities introduced by non- 
head nouns in a compound nominal. We refer the 
reader to (Vieira, 1998) for discussion of the heuris- 
tics for this last class. 
Our system attempts to resolve lexical bridges by 
consulting WordNet to determine if there is a se- 
mantic relation between the head noun of the de- 
scription and the head noun of one of the NI'S in the 
previous five sentences. The results of this search 
for our training corpus, in which 204 descriptions 
are classified as bridging, are shown in Table 4. It is 
interesting to note that the semantic relations found 
in this automatic semch were not always those ob- 
served in our manual analysis. 
Bridging 
Class 
Synonimy 
Hyponimy 
Meronimy 
Sister 
Relations 
Found 
11 
59 
30 
Total 106 30 
Right % 
Anchors Right 
4 36% 
18 30% 
2 33% 
6 20% 
28% 
Table 4: Ewduation of the search for anchors using 
WordNet 
We developed a simple heuristic method for as- 
signing types to named entities. Our method iden- 
titied entity types for 66% (535/814) of all names 
in the corpus (organizations, persons and locations). 
The precision was 95%. We could have had a better 
recall if we had adopted more comprehensive lists 
of cue words, or consulted dictionaries of names 
as done for the systems participating in MUC-6. 
There, recall in the named entity task varies t'rom 
82% to 96%, and precision l'rom 89% to 97%. 5 
4 Overall Evaluation of the System 
The order of application of heuristics is as impof 
tant as the heuristics themselves. The final order of 
application was also arrived at on the basis of an ex- 
tensive evaluation (Vieira, 1998), and is based on 
the l'ollowing strategy: 6 
5A more recent version o1' the system using the named en- 
tity recognition software developed by ItCRC for the MUC-7 
competition (Mikheev el al., 1999) is discussed in (Isbikawa, 
1998). 
aWe also attempted to learn the best order of application of 
the heuristics automatically by means of decision tree learn- 
ing algorithms (Quinlan, 1993), without however observing a 
signflicant difference in pcrfommnce. See (Vieira, 1998) for 
details. 

References

D. et al.. Appelt. 1995. SRI International FASTUS system
MUC-6 test results and analysis. In Proc. of
the Sixth Message Understanding Conference, pages
237-248, Columbia, Maryland, November.

D. L. Bean and E. Riloff. 1999. Corpus-based identification
of non-anaphoric noun phrases. In Proc. of the
37th ACL, pages 373-380, University of Maryland.

J. Carletta. 1996. Assessing agreement on classification
tasks: the kappa statistic. Computational Linguistics,
22(2):249-254.

D. M. Carter. 1987. Interpreting Anaphors in Natural
Language Texts. Ellis Horwood, Chichester, UK.

H. H. Clark and C. R. Marshall. 1981. Definite reference
and mutual knowledge. In A. Joshi, B. Webber, and
I. Sag, editors, Elements of Discourse Understanding.
Cambridge University Press, New York.

H. H. Clark. 1977. Inferences in comprehension. In
D. Laberge and S. J. Samuels, editors, Basic Process
in Reading: Perception and Comprehension.
Lawrence Erlbaum.

C. Fellbaum, editor. 1998. WordNet: An electronic lexical
database. The MIT Press.

B. A. Fox. 1987. Discourse Structure and Anaphora.
Cambridge University Press, Cambridge, UK.

K. Fraurud. 1990. Definiteness and the processing
of NPs in natural discourse. Journal of Semantics,
7:395-433.

B. J. Grosz and C. L. Sidner. 1986. Attention, intention,
and the structure of discourse. Computational
Linguistics, 12(3):175-204.

J. A. Hawkins. 1978. Definiteness and Indefiniteness.
Croom Helm, London.

M. A. Hearst. 1997. TextTiling: Segmenting text into
multi-paragraph subtopic passages. Computational
Linguistics, 23(1):33-64.

K. Humphreys, R. Gaizauskas, S. Azzam, C. Huyck,
B. Mitchell, and H. Cunningham Y. Wilks. 1998.
University of Sheffield: description of the LaSIE-II
System as used for MUC-7. In Proc. of the Seventh
Message Understanding Conference (MUC-7).

T. Ishikawa. 1998. Acquisition of associative information
and resolution of bridging descriptions. Master�s
thesis, University of Edinburgh, Department of Linguistics,
Edinburgh, Scotland.

S. Loebner. 1987. Natural  and generalised
quantifier theory. In P. G�ardenfors, editor, Generalized
Quantifiers, pages 93-108. D. Reidel, Dordrecht.

D. Marcu. 1999. A decision-based approach to rhetorical
parsing. In Proc. of the 37th ACL, pages 365-372,
University of Maryland, June. ACL.

A. Mikheev, M. Moens, and C. Grover. 1999. Named
Entity recognition without gazetteers. In Proc. of
EACL, Bergen, Norway. EACL.

M. Poesio and R. Vieira. 1998. A corpus-based investigation
of definite description use. Computational Linguistics,
24(2):183-216, June.

M. Poesio, S. Schulte im Walde, and C. Brew. 1998.
Lexical clustering and definite description interpretation.
In Proc. of the AAAI Spring Symposium on
Learning for Discourse, pages 82-89, Stanford, CA,
March. AAAI.

E. F. Prince. 1981. Toward a taxonomy of given-new
information. In P. Cole, editor, Radical Pragmatics,
pages 223-256. Academic Press, New York.

J. R. Quinlan. 1993. C4.5: programs for machine learning.
Morgan Kaufmann, San Mateo, CA.

R. Reichman. 1985. Getting Computers to Talk Like You
and Me. The MIT Press, Cambridge, MA.

C. L. Sidner. 1979. Towards a computational theory
of definite anaphora comprehension in English discourse.
Ph.D. thesis, MIT.

K. Strand. 1996. A taxonomy of linking relations.
Manuscript. A preliminary version presented at the
Workshop on Indirect Anaphora, Lancaster University,
1996.

R. Vieira and M. Poesio. 1999. Processing definite
descriptions in corpora. In S. Botley and
T. McEnery, editors, Corpus-based and Computational
Approaches to Discourse Anaphora. John Benjamins.

R. Vieira. 1998. Definite Description Resolution in Unrestricted
Texts. Ph.D. thesis, University of Edinburgh,
Centre for Cognitive Science, February.
