Abstract Anaphora Resolution in Danish 
Costanza Navarretta 
Center for Sprogteknologi 
Njalsgade 80 
2300 Kcbenhavn S 
costanza@cst.ku.dk 
Abstract 
In this paper 1 I describe the use 
of Danish pronouns and deictics 
in dialogues. Then I present an 
adaptation to Danish of Eckert 
and Strube's algorithm for resolv- 
ing anaphora referring to individ- 
ual NPs and abstract objects in En- 
glish dialogues (Eckert and Strube, 
1999b; Eckert and Strube, 1999a). 
The adapted algorithm is tested on 
four Danish dialogues from two dia- 
logue collections and the results ob- 
tained are evaluated. 
.1 Introduction 
Many natural  processing applica- 
tions involve the complex task of resolving 
anaphora. Different strategies for anaphora 
resolution have been proposed, some exclu- 
sively relying on the syntactic structure of 
discourse, some including semantic and prag- 
matic constraints, some based on statisti- 
cal methods. One of the most popular ap- 
proaches to anaphora resolution is center- 
ing (Grosz et al., 1995), henceforth GJW95, 
which accounts for the relation between the 
saliency of entities in discourse and the use 
of referring expressions, incorporating syn- 
tax, semantics and pragmatics. Centering fits 
into Grosz and Sidner's model of discourse 
structure (Grosz and Sidner, 1986). In this 
model a discourse is composed of segments 
which exhibit global coherence. A discourse 
1This work has been carried out under Staging, an on-going Danish project funded by the Danish Re- 
search Councils. 
segment, on the other hand, is composed of 
a sequence of utterances which exhibit local 
coherence. This latter phenomenon is ac- 
counted for by centering theory. Centering 
predicts that there is a connection between 
the coherence of a referring expression and 
the inference load necessary to resolve it. Al- 
though Grosz, Joshi and Weinstein recognize 
that many factors determine the prominence 
of entities in an utterance, in GJW95 this 
prominence is established simply by the linear 
order of the entities in the utterance. Differ- 
ent centering algorithms have been presented, 
spelling out the strategy described in GJW95, 
extending the theory to more linguistic phe- 
nomena or specifying the concept of promi- 
nence of discourse entities. Strube and Hahn 
(Strube, 1998; Strube and H~.hn, 1999) in par- 
ticular, calculate prominence considering the 
information structure of the utterances (func- 
tional centering). 2 The prominence ranking 
they adopt does not exclusively rely on word 
order, which is  dependent. More- 
over GJW95 only dealt with intersentential 
anaphora, while Strnbe and Hahn account 
for both intrasentential and intersentential, 
pronominal and nominal anaphora, a 
Centering-based algorithms have been 
tested on written texts. Recently they have 
also been applied to written dialogues. Byron 
and Stent (1998), in particular, test center- 
ing on multi-party dialogues. They conclude 
that centering seems to be a valid theory also 
in this case, but it must be extended to ac- 
2In (Strube and Hah~ 1996) a functional-based 
prominence ranking has been proposed. 
SAn other extension of the centering framework 
to intrasentential anaphora has been proposed by 
Kameyama (1998). 
56 
count for dialogne-specific aspects such as the 
definition of utterance boundaries, the speci- 
fication of a strategy for tackling partial ut- 
terances and including discourse participants 
in the list of relevant discourse entities. 
Eckert and Strube (1999a; 1999b), hence- 
forth ES99, describe an algorithm for resolv- 
ing anaphors referring to individual NPs and 
abstract objects in English dialogues. The al- 
gorithm is based on rules for discriminating 
among the two types of anaphor based on the 
predicative contexts in which the anaphors 
occur. The individual anaphors are then re- 
solved by the functional centering algorithm 
described in (Strube, 1998), while abstract 
anaphors are resolved with a different algo- 
rithm. ES99 test the approach on selected 
dialogues and obtain a precision of 63,6% 
for discourse deictics and 66,2% for individ- 
ual anaphors. They report that most errors 
are due to the inability to distinguish be- 
tween discourse deictics and pronouns which 
vaguely refer to concepts in the preceding dis- 
course (vague anaphors). Another cause of er- 
ror is the lack of information about abstract 
nominals. I believe that the strategy followed 
by ES99 is a good starting point for inves- 
tigating how far one can go in resolving in- 
dividual and abstract anaphors in dialogues 
on the basis of the local contexts in which the 
anaphors occur. I have adapted the algorithm 
so it accounts for Danish data and have ap- 
plied it to Danish dialogues. 4 
In section 2 I shortly present the original 
centering framework and functional centering 
as described in (Strube, 1998), $98. In sec- 
tion 3 Eckert and Strube's algorithm is in- 
troduced and in 4 the Danish personal and 
demonstrative prononn~ are described with 
focus on discourse deictics in dialogues. In 
section 5 1 present my adaptation of the ES99- 
algorithm to Danish data. Section 6 con- 
talus an evaluation of the results obtained by 
manually testing the adapted ES99-algorithm 
on randomly selected dialogues from the col- 
lection "Samtale hos Leegen" (Conversation 
at the doctor's) (SL) and "ProjektIndvaudr- 
4Centering-based algorithms have recently been 
tested on Danish discourse (Navarretta, 2000). 
erdansk" (Project Immigrant Danish) (PID), 
collected by researchers at the Department of 
General and Applied Linguistics of the Uni- 
versity of Copenhagen. In section 7 I outline 
future work for improving the results of the al- 
gorithm and make some concluding remarks. 
2 Centering 
In GJW95 the entities which link an utterance 
Un to the others in the same discourse seg- 
ment are the centers of that utterance. Each 
utterance is assigned a set of forward-looking 
centers, Cf, and, with the exception of the 
initial utterance of the segment, a backward- 
looking center, Cb. The Cb of an utterance Un 
connects with one of the forward-looking cen- 
ters of the preceding utterance Un-1 while the 
forward-looking centers only depend on the 
expressions in Un. The forward-looking cen- 
ters are partially ordered to reflect relative 
prominence. GJW95 recognize three types of 
transition relation across pairs of utterances: 
continue, retain and shift (see table 1). 
Center movement and realization are con- 
strained by two rules: 
Rule I: If any element of CCf(U~-i) is real- 
ized by a pronoun in Un, then Cb(Un) 
must also be realized by a pronoun 
Rule 2: Center continuation is preferred to 
center retaining which is preferred to cen- 
ter shifting 
2.1 Functional Centering 
In $98 the functions of the backward-looking 
center and the transitions in the centering 
theory are replaced by the order of elements 
in a list of salient discourse entities, the S- 
list. The ranking criteria for the elements in 
the S-list are based on (Prince, 1981), where 
discourse entities are classified into hearer- 
old (OLD), mediated (MED) and hearer-new 
(NEW). The two tuples (x, Uttx, posx) and 
(y, utty, posy) in the S-list indicate that the 
entity x is evoked in utterance uttx at posi- 
tion posx and that y is evoked in utterance 
utty at position posy respectively. Given that 
Uttx and utty refer to Un or Un-1, the follow- 
57 
Table 1: Transition States 
OR no Cb(U.-1) 
CONTINUE 
RETAIN 
SHIFT 
ing ranking constraints on the S-list entities 
are valid (Strube, 1998)\[p.1253\]: s 
1. ifxEOLDandyEMED, thenx~y 
if x E OLD and y E NEW, then x --< y 
ifx E MED and y E NEW, then x ~ y 
2. if x,y E OLD or x,y E MED or x,y E 
NEW, 
then if uttx > Utty then x ~ y 
if utt~ = utty and posz < posy then x -~ y 
The S98-algorithm Consists in testing a re- 
ferring expression against the elements in the 
S-list from left to right until the test suc- 
ceeds. The S-list is then updated so that new 
elements are inserted according to the S-list 
ranking criteria. When the analysis of an ut- 
terance is finished all the entities which were 
not realized in the utterance axe removed from 
the S-list. 
3 Eckert and Strube's Algorithm 
ES99 propose a new algorithm for resolving 
anaphors with abstract object antecedents. 
Analyzing a collection of telephone conversa- 
tions they distinguish the following anaphor 
types: individual anaphors, discourse deic- 
tics, inferrable-evoked anaphors 6 and vague 
anaphors. Other types of pronoun are not 
taken into consideration. 
Predicates that are preferentially associ- 
ated with abstract objects are marked as I- 
incompatible (*I) while predicates that are 
preferentially associated with individual ob- 
jects are marked as A-incompatible (*A). 
5I mark ranking precedence with ~. 
61nferrable-evoked anaphors refe~ to the use of the 
plttral pronoun they indirectly co-specifying with a 
singular NP which indicates a country or an institu- 
tion. 
ES99 define the following *I predicates 
(Eckert and Strube, 1999b)\[p. 40\]: 
Equating constructions where a pronom- 
inal referent is equated with an abstract 
object, e.g., x is making it easy, x is a 
suggestion. 
Copula constructions whose adjectives 
can only be applied to abstract entities, 
e.g., x is true, x ks false, x is correct, x is 
right, x isn't right. 
Arguments of verbs describing 
propositional attitude which take 
S'-complements, e.g., assume. 
Object of do. 
Predicate or anaphoric referent is a "rea- 
son", e.g., x is because I like her, x is why 
he's late. 
Predicates that are preferentially associ- 
ated with individual objects are the following 
(Eckert and Strube, 1999b)\[p. 40\]: 
Equating constructions where a pronom- 
inal referent is equated with a concrete 
individual referent, e.g., x is a ear. 
Copula constructions with adjectives 
which can only be applied to concrete en- 
tities, e.g., x is expensive, x is tasty, x is 
loud. 
Arguments of verbs describing physi- 
cal contact/stimulation, which cannot be 
used anaphorically, e.g., break x, smash z, 
eat x, drink x, smell x but NOT *see x 
Grounded acts are used as domain for the 
anaphor resolution algorithms in dialogues. 
58 
In particular two dialogue acts, Initiations 
(Is) and Acknowledgments (As) are rec- 
ognized. Is have semantic content, while As 
are only used to ground the preceding I. Ac- 
knowledgments/Initiations (A/Is) are di- 
alogue acts that have both the function of 
grounding the preceding I and that of estab- 
lishing a new I. An I and the correspond- 
ing A, together with longer Is in the same 
turn-taking which do not need to be acknowl- 
edged, constitute a Synchronizing Unit 
(SU). Short Is which are not acknowledged 
are ignored by the resolution algorithms. 
ES99 follow i.a. (Webber, 1991) in assum- 
ing that anaphoric discourse deictic reference 
involves reference coercion and that only dis- 
course sections adjacent to the anaphor or, 
using Webber's terminology, sections on the 
right frontier of the discourse structure tree, 
are available for discourse-deictic reference. 
Like (Asher, 1993) they assume that the type 
of abstract object is determined by the con- 
text in which the anaphor occurs. Anaphora 
referring to abstract objects are resolved us- 
ing a list, the A-list. The A-list is only filled 
when discourse deictics occur and its elements 
remain for one I. The parts of the linguistic 
contexts are accessed in the following order: 
1. the A-list; 2. in the same I the clause to the 
left of the clause which contains the anaphor; 
3. within the previous I the rightmost main 
clause and subordinated clauses to its right; 
4. within previous Is the rightmost complete 
sentence, if previous I is an incomplete sen- 
tence. 
The anaphora resolution algorithm for 
third person singular neuter personal pro- 
nouns is the following (Eckert and Strube, 
1999a): 
case PRO is I-incompatible 
if resolveDiscourseDeictic(P RO ) 
then classify as discourse deictic 
else classify as vague pronoun; 
case PRO is A-incompatible 
if resolveIndividual (PRO ) 
then classify as individual pronoun 
else classify as vague pronoun; 
case PRO is ambiguous 
if resolveIndividual(PRO ) 
then classify as individual pronoun 
else if resolveDiscourseDeictic(PRO) 
then classify as discourse deictic 
else classify as vague pronoun; 
The same algorithm is used for demon- 
stratives, with the exception that the last 
two if constructions in the algorithm for pro- 
nouns are reversed reflecting the preference 
for demonstratives to be discourse deictics 
(Webber, 1991). 
4 Danish Data 
In this section I shortly describe Danish third 
person personal and possessive pronouns and 
demonstrative pronouns. The description fo- 
cuses on the discourse deictic use of these pro- 
nouns based on occurrences in three Danish 
dialogue collections, Bysoc, 7 SL and PID. 
My description is also based on (Allan et al., 
1995). The third person singular personal and 
possessive pronouns can be found in table 2, 
while the third person plural personal and 
possessive pronouns can be found in table 3. s 
Den, det and de are also used as defi- 
nite articles (the) and demonstrative deter- 
miners (this/that and these/those). In spo- 
ken  the demonstratives are always 
stressed. 9 Den, det, de are demonstratives 
if followed by the adverbials her and der in 
which case they correspond to the English 
this/these and that/those respectively. Fur- 
thermore, the demonstratives denne, dette 
(this) and disse (these) exist. 
Femjnlne and masculine pronouns generally 
co-refer 1° with persons, but can also refer to 
pets as in English. Common gender pronouns 
refer to common gender nouns which do not 
denote humans. Common gender nouns de- 
noting humans are neutral as to the sex of 
the person they refer to. Thus the gender of 
7The Bysoc corpus has been collected by re- 
searchers at Copenhagen University under "Projekt 
Bysociolingvistik" (Project Urban Sociolinguistics). 
SThe Danish reflexive pronouns are used differently 
than the English ones, see i.a. (Neville, 1998). 
9Because I do not have access to phonetic infor- 
mation about the considered dialogues I cannot ac- 
count for important phenomena such as intonation 
and prosody, see i.a. (Vallduv~ and Engdahl, 1995). 
1°From now on I will simply write "refer to". 
59 
Table 2: Third person singular pronouns 
gender subject object reflexive possessive pos.refl. 
feminine hun she hende her sig herself hendes hers si-nJtJne hers 
masculine 
COmmOn 
neuter 
han he 
den it 
det it 
ham him sig himself 
den it sig itself 
det it sig itself 
hans his si-nftfne his 
dens its si-n/tfne its 
dets its si-n/t/ne its 
Table 3: Third person plural pronouns Lsubject I °b Jeer Lre exive IP° I 
de they dem them sig themselves deres their/theirs 
the referring pronoun corresponds to the sex 
of the person the noun refers to. 
Neuter gender pronouns are used to refer 
to neuter nouns. They can also refer to a 
few common person nouns in neuter gender, 
such as barn (child) and menneske (person) 
if the sex of the person is unknown or irrel- 
evant (syntactic agreement). In case the sex 
is known or relevant, the appropriate femi- 
nine or masculine pronouns are used (seman- 
tic agreement). The two cases are illustrated 
in the following examples: 
barnet var pd millimeter sd stort 
det skulle v,~re i l,~ngden og i hov- 
edstcrrelsen og... 
(the child was precisely as high as it 
ought to be and its head was as big 
as it ought to and... ) 
sd ch... jeg kunne gd ud \]or jeg 
havde mit barnebarn reed pd tre et 
halvt dr sd..., kunne jeg jo bare holde 
hami hdnden 
(so oh... I could leave because I was 
together with my three and half year 
old grandchild so..., I could just hold 
his hand) 
Both den and det can refer to collective 
nouns. In this case the choice between the 
singular den or det and plural de depends on 
whether the speaker focuses on the collective 
meaning or on the individuals. Det and in 
few idiomatic expressions den are also used 
as expletives. 
In Danish the most frequently used dis- 
course deictic is det which corresponds to it, 
this or that. Other discourse deictics are det 
her (this) and det der (that). These two deic- 
tics can be used in most of the same contexts 
as det, although there seems to be a prefer- 
ence for using them to refer to several clauses. 
The neuter demonstrative dette (this) has also 
a discourse deictic use, but is mostly used in 
written . I did not found any occur- 
rences of it in the three dialogue collections. 
As discourse deictic det refers to an infini- 
tive or a clause, as it is the case in the follow- 
ing examples: 
At ryge er \]arligt og det er ogsd dyrt 
(Smoking is dangerous and it is also 
expensive) 
A: Du skal rage en blodprcve 
(You have to take a blood test) 
B: Hvorffor det? 
(Why that.*) 
Det is also used as the subject complement 
of vmre (be) and blive (become) in answers. 
A: Blev du ff,~rdig reed opgaven? 
(Were you done with the task?) 
B: Ja, det blev jeg 
(lit. Yes, that was I) 
(Yes, I was) 
Det refers to a verb phrase when it is used 
as the object complement for the verb have 
(have), gCre (do) and modal verbs as in 
60 
Alle faldt, men det gjorde jeg ikke 
(lit. All fell, but that did I not) 
(All fell, but I did not) 
Det refers to a clause in constructions with 
attitude verbs and other verbs which take 
clausal complements, such as synes (think), 
fro (believe) and vide (know), sige (say), hdbe 
(hope): 
A: Det begynder snart at regne. 
(It will soon begin to rain) 
B: Det hdber jeg ikke 
(lit. That hope I not) 
(I hope not) 
In the latter three cases the pronoun det is 
often topicalized, i.e. it appears before the 
main verb, in the place that usually is occu- 
pied by the subject 11. 
Det can also refer to more clauses, or to 
something that can vaguely be inferred from 
the discourse. 
A: barnets .far chin ... 
(the baby's father uhm ...) 
B: ja 
(yes) 
A: havde alvorlig, 
ch. . . spmdbcrnsgulsot da han 
blev fcdt 
(had serious, uh ... infant icterus 
when he was born) 
B: ja 
(yes) 
A: og fik sd ogs~ skirter sit blod ikke 
ogs~ 
(and then he also got a blood 
transfusion, didn't he) 
B: mmh 
A: det havde hans storebror ogs~i 
(lit. that had his brother too) 
(his brother had it too) 
B: ja 
(yes) 
A: og er blevet hjerneskadet afdet 
(and he got brain damage from it) 
B: ja 
11This position is called fundamentfelt (actualiza- tion field) by (Diderichsen, 1984 1946). 
(yes) 
A: altsd jeg red ikke om deter noget 
jeg skal, om deL skal skrives nogen 
steder eller gCres noget red 
(so I don't know whether it is 
something I should do, whether it 
should be written somewhere or 
something should be done) 
In the above example the deictics in the 
last utterance do not refer to a single clause 
or predicate, but to the whole family history 
of icterus. 
To conclude, Danish deictics are used in 
more contexts than the English ones. Espe- 
cially noticeable is the Danish use of discourse 
deictics in cases where elliptical constructions 
are normal in English. 12 
5 The Adapted ES99-algorithm 
On the basis of the deictics in the two Danish 
dialogue corpora, SL and PID I have estab- 
lished the following *I predicates for Danish: 
constructions where a pronoun is equated 
with an abstract object, e.g., x er et 
forslag (x is a suggestion) 
copula constructions with adjectives 
which can only be applied to abstract en- 
tities, such as x er sandt (x is true), x er 
usandt (x is untrue), x er rigtigt (x is 
correct) 
arguments of verbs which take S'- 
complements, e.g., fro (believe), antage 
(ass-me), mene (think), sige (say) 
anaphoric referent in constructions such 
as x er /ordi du er holdt op reed at ryge 
(x is because you have stopped smoking) 
x er pd grund af at duer gravid (x is 
because you are pregnant) 
• object of g#re (do) 
* subject complement with vmre (be) and 
blive (become) in answers 
12I have not included in the description cataphoric 
deictic pronouns. 
61 
• object of have (have) if" the verb was not 
used as a main verb in the previous clause 
• object of modal verbs 
The last four predicates are specific for 
Danish. I have assumed the following *A 
predicates, which are mainly' translations of 
the English ones: 
• constructions where a pronominal refer- 
ent is equated with a concrete individual 
referent, such as x er en legemsdel (x is 
a body part), x er et barn (x is a baby) 
• copula constructions with adjectives 
which can only be applied to concrete en- 
tities, such as x er rcdt (x is red) 
• arguments of verbs describing physical 
contact/stimulation, which cannot be 
used anaphorically, e.g. spise x (eat x), 
drikke x (drink x) 
As Eckert and Strube notice for English, 
also in Danish there are cases where the con- 
texts of an anaphor can allow both an in- 
dividual NP and an abstract object. Some 
examples are copula constructions like x er 
godt/ddrligt (x is good/bad), and objects of 
verbs such as elske (love), hade (hate), fore- 
traekke (prefer). To partially accomodate this, 
I have added the following condition to the 
algorithm: in the above cases the anaphor is 
classified as A* incompatible unless the pre- 
vious clause contain.~ a raising adjective con- 
struction in which case it is considered I* in- 
compatible. Consider the fi)llowing two ex- 
amples: 
Peter boede iet r~dt hus. Det 
hadede han. 
(Peter lived in a red house. He 
hated it.) 
Deter dcdsygt at sidde pd et vaskeri. 
Det hader jeg. 
(It is boring to be in a laundry. I 
hate it) 
In the first example the algorithm chooses 
et r~dt hus (a red house) as the antecedent 
of det, while in the second example the algo- 
rithm chooses at sidde pd et vaskeri (being in 
a laundry) instead of et vaskeri. There are 
cases, similar to the first example, where it is 
impossible, without a deeper analysis of the 
discourse to determine whether an anaphor 
refers to an individual NP or an abstract ob- 
ject. 
In the test I have taken into account the 
metaphorical uses of verbs encoded in a se- 
mantic lexicon, the Danish SIMPLE lexicon 
(Pedersen and Nimb, 2000). 
From the analysis of anaphors in the con- 
sidered dialogue collections I found that many 
individual anaphors refer back to entities 
which have not been evoked in the immedi- 
ately preceding utterances (SUs) and thus 
they would not be on the S-list (the enti- 
ties which are not evoked in the current SU 
are removed from the list). Thus I have ex- 
tended the scope of resolution for all individ- 
ual anaphors except the neutral singular. If 
an antecedent to an individual NP cannot be 
resolved by looking at the actual S-list, the 
elements on the S-lists for the preceding SUs 
are considered. 13 
6 Evaluation of the Algorithm 
I have applied the modified ES99-algorithm to 
three randomly selected SL dialogues (6,305 
words) and to one of the dialogues between 
native Danes recorded in the PID collection 
(5,367 words). It must be noted that in my 
test only one annotator (the author) iden- 
tiffed dialogue acts, classified the anaphors 
in the dialogues, marked NPs and anaphor 
antecedents. In (Eckert and Strube, 1999a) 
these tasks have been accomplished by two 
annotators. 
In dividing the three SL dialogues into dis- 
course segments I have mainly used a parti- 
tion made by two researchers at the Univer- 
sity of Copenhagen in an independent project. 
The discrimination criteria were topic shift 
and a few linguistic clues. I have then ap- 
IsI have followed the cache model described in 
(Walker, 1998). In the present test it was necessary to 
go back maximally seven SUs to find an antecedent 
to an individual pronominal anaphor. 
62 
plied the same discrimination criteria to the 
dialogue from the PID collection. 
I have defined dialogue units syntactically 
following (Eckert and Strube, 1999a). 14 
Because it is not always possible to dis- 
tinguish between den, det, de used as per- 
sonal or demonstrative pronouns without hav- 
ing access to stress information, I have classi- 
fied them as personal pronouns unless they 
are topicalized, or occur in syntactic con- 
structions where demonstratives are normally 
used. The manual classification of pronouns 
and demonstratives in the four dialogues can 
be found in table 4. 
The results of the individual anaphora reso- 
lution algorithm can be found in table 5, while 
the results of the discourse deictics resolution 
algorithm are given in table 6. 
The results obtained are better than those 
reposed in (Eckert and Strube, 1999a), but I 
have used more background information than 
ES99 and extended the scope of resolution for 
individual anaphors (without this extension 
the precision of the individual resolution al- 
gorithm was of 64.5). Furthermore the Dan- 
ish deictic det occurs in more contexts than 
the English it, this and that, thus there are 
more I* predicates in the Danish version of 
the algorithm than in the original one. The 
fact that only one annotator divided the dia- 
logues into SUs may also have influenced the 
results. 
The algorithm classifies anaphors and re- 
solves some of them, thus there are two types 
of error, classification errors and resolution er- 
rors. Most of the instances of wrongly classi- 
fied anaphors are due to the fact that the al- 
gorithm classifies vague anaphors as discourse 
deictics and then resolves the anaphor to a 
preceding predicate or clause. Few errors are 
due to the fact, already noticed by ES99, that 
the defined I* and A* predicates do not con- 
tain information about nominals referring to 
abstract objects. 15 These errors resulted in 
most cases in resolution errors. 
Some errors are due to the inability to find 
14The dialogue collections have been tagged. 
15The semantic lexicon I used did not contain the 
relevant nominals. 
an individual NP antecedent to the pronoun 
det, when this refers generally to an NP of dif- 
ferent gender 16 and to wrongly resolved plu- 
ral pronouns with complex NP antecedents or 
with no antecedent. Correctly classified, but 
wrongly resolved discourse deictics are, i.a., 
due to the fact that I did not mark in any 
particular way parenthetical utterances. The 
latter kind of errors are chaining errors. In 
table 7 the occurrences of each type of error 
are reported. 
7 Concluding Remarks 
The adapted ES99-algorithm has been tested 
on two kinds of dialogue, that have been clas- 
sifted by one annotator. Although the types 
of dialogue in the Danish test is quite different 
from that used by ES99, the results reported 
in the previous section (6) indicate that the 
algorithm performs as well for Danish as for 
Enghsh. Because the use of Danish pronouns, 
especially those referring to abstract objects, 
is different from the English one, these results 
provide an interesting evaluation of the algo- 
rithm. 
As noticed by ES99, adding more lexical 
knowledge to the algorithm could improve its 
performance. I also beheve that the con- 
texts of abstract anaphors should be studied 
in more dialogues, and that more attention 
should be given to the connection between 
discourse deictics and the relations that \]in~ 
pieces of discourse to each other (Webber, 
1991; Fraurud, 1992; Asher, 1993; Kehler, 
1997). 
Further work will thus consist in analyzing 
the occurrences of discourse deictics in both 
written texts and dialogues and paying 
additional attention to the relations linking 
pieces of discourse to each other (i.a. (Hobbs, 
1979; Mann and Thompson, 1987; Polanyi, 
1988) 
leThe use of especially generic plural pronouns in 
Swedish is discussed in (Fraurud, 1992). 
63 
Table 4: Classification of Pronouns and Demonstratives 
A5 AA10 AAll TR3' ~ 
Individual Pro 39 43 34 51 167 
Discourse Deictics Pro 25 16 17 34 92 
Vague Pro 4 6 0 2 12 
Inferrable Evoked 1 0 0 2 3 
Individual Dem 1 5 0 4 10 
Discourse Deictics Dem 27 20 19 28 94 
Vague Dem 2 5 3 3 13 
Table 5: Results of the Individual Anaphora Resolution Algorithm 
A5 AA10 AAll TR1 
No. resolved correctly 31 40 23 40 134 
No. of Individual Pro 41 48 34 57 I 180 
Precision 0.756 0.833 0.676 0.701~0.744 
Table 6: Results of the Discourse Deictics Resolution Algorithm 
No. resolved correctly 
No. of Discourse Deictics 
Precision 
A5 AA10 AAll TR1 
43 25 33 49 147 
58 47 39 67 211 
0.741 0.489 0.846 0.716 0.696 
Table 7: Wrongly resolved anaphors 
DD-vague 10 
wrongly resolved plural 13 
generic det 7 
abstract nomina|~ 5 
individual anaphora instead of DD (% abstract nominals) 6 
~rrong!y resolved DD 28 
wrongly resolved individual anaphora (singular) 2 
chaining errors 39 
64 

References 
Robin Allan, Philip Holmes, and Tom Lundsk~er- 
Nielsen. 1995. Danish - A Comprehensive 
Grammar. Routledge, London. 
Nicholas Asher. 1993. Reference to Abstract 
Objects in Discourse, volume 50 of Studies in 
Linguistics and Philosophy. Kluwer Academic 
Publishers, Dordrecht, the Netherlands. 
D. Byron and A. Stent. 1998. A Preliminary 
Model of Centering in Dialog. In Proceedings 
of the 36th Annual Meeting, volume II, pages 
1475-1477. Association for Computational Lin- 
guistics. 
Paul Diderichsen. 1984-1946. Elementa;r Dansk 
Grammatik. Gyldendal, Copenhagen. 
Miriam Eckert and Michael Strube. 1999a. Dia- 
logue Acts, Syncchronising Units and Anaphora 
Resolution. In Jan van Kuppevelt, Robert van 
Leusen, Robert van Rooy, and Henk Zeevat, 
editors , Amstelogue '99 Proceedings - Workshop 
on the Semantics and Pragmatics of Dialogue. 
Miriam Eckert and Michael Strube. 1999b. Re- 
solving Discourse Deictic Anaphora in Dia- 
logues. In Proceedings of the 9th Conference 
of the European Chapter of the Association of 
Computational Linguistics, pages 37--44. 
Kari Fraurud. 1992. Processing Noun Phrases in 
Natural Discourse. Department of Linguistics - 
" Stockholm University. 
Barbara Grosz and Candace Sidner. 1986. Atten- 
tion, Intentions, and the Structure of Discourse. 
Computational Linguistics, 12(3):175-284. 
Barbara Grosz, Aravind K. Joshi, and Scott Wein- 
stein. 1995. Centering:A Framework for Mod- 
eling the Local Coherence of Discourse. Com- 
putational L inguistics~ 21(2):203-225. 
Jerry R. Hobbs. 1979. Coherence and Corefer- 
ence. Cognitive Science, 3(1):67-90. 
Megumi Kameyama. 1998. Intrasentential cen- 
tering: A case study. In Marilyn A. Walker, 
Aravind K. Joshi, and Ellen F Prince, editors, 
Centering Theory in Discourse, pages 89--112. 
Oxford University Press. 
Andrew Kehler. 1997. Current Theories of 
Centering for Pronoun Interpretation: A crit- 
ical Evaluation. Computational Linguistics, 
23 (3) :467-475. 
William C. Mann and Sandra A. Thompson. 
1987. Rhetorical Structure Theory: De- 
scription and Construction of Text Structures. 
In Gerard Kempen, editor, Natural Language 
Generation, number 135 in NATO ASI, pages 
85-95. Martinus Nijhoff Publishers. 
Costanza Navarretta. 2000. Centering-based 
Anaphora Resolution in Danish. In Petr Sojka, 
Ivan Kopecek, and Karel Pala, editors, Pro- 
ceedings of the Third International Workshop 
on Text, Speech, Dialogue - TSD 2000, pages 
345-350, Brno, Czech Republic. 
Anne Neville. 1998. Subject-orientedness and Ex- 
empt Anaphors. In Proceedings of the Xth Con- 
ference of Nordic and General Linguistics Uni- 
versity of Iceland, Reykjavik. 
Bolette S. Pedersen and Sanni Nimb. 2000. Se- 
mantic Encoding of Danish Verbs in SIMPLE - 
Adapting a verb-framed model to a satellite- 
framed . In Proceeding from Sec- 
ond International Conference on Language Re- 
sources and Evaluation - LREC 2000, Athens. 
Livia Polanyi. 1988. A Formal Model of the 
Structure of Discourse. Journal of Pragmatics, 
12(5/6):601-638. 
Ellen F. Prince. 1981. Toward a taxonomy of 
given-new information. In Peter Cole, editor, 
Radical Pragmatics, pages 223-255. Academic 
Press. 
Michael Strube and Udo Hahn. 1996. Functional 
Centering. In Proceedings of the 34th Annual 
Meeting, pages 270-277. Association for Com- 
putational Linguistics. 
Michael Strube and Udo Hahn. 1999. Functional 
Centering - Grounding Referential Coherence 
in Information Structure. Computational Lin- 
guistics, 25(3):309-344. 
Michael Strube. 1998. Never Look Back: An Al- 
ternative to Centering. In Praeeedings of the 
36th Annual Meeting, volume II, pages 1251- 
1257. Association for Computational Linguis- 
tics. 
Enric Vallduv~ and Elisabet Engdahl. 1995. The 
linguistic realization of information packaging. 
Linguistics, 34(3):459-519. 
Marilyn A. Walker. 1998. Centering, Anaphora 
Resolution, and Discourse Structure. In Mari- 
lyn A. Walker, Aravind K. Joshi, and Ellen F. 
Prince, editors, Centering Theory in Discourse. 
Oxford University Press. 
Bonnie L. Webber. 1991. Structure and 
Ostension in the Interpretation of Discourse 
Deixis. Natural Language and Cognitive Pro- 
cesses, 6(2):107-135. 
