Discourse Anaphora 
Joke Dorrepaal 
OTS RUU Utrecht 
dorrepaal@hutruu59.BITnet 
1 .Abstract 
This paper reports on a model that serves 
anaphora resolution. A distinction will be made 
between possible antecedents and preferred an.- 
tecedents. The set of linguistically possible can-- 
didates will be detined in terms of compatibility 
and recency. Preferred antecedents are a sub- 
set of the possible antecedents, selected by the 
application of extralinguistic knowledge. Moti- 
vation for the particular design and comparison 
with other approaches are extensive. 
2 Introduction 
Both in formM semantics and in NLP, the issue 
of (pro)nominal reference has aroused much in- 
terest. Formal discourse semantics was mainly 
inspired by the Discourse Representation The- 
ory developed by Kamp (1981) and, the closely 
related, File. Change Semantics (tIeim, 1982). In 
standard I)R;I', hardly any restrictions were im- 
posed on the relation between antecedent and 
anaphor. 
This contrasts strongly with work done in the 
NLP area. Knowledge of the world and intricate 
inferences play key roles in restricting the num- 
ber of antecedents for anaphora. Insights from 
artificial intelligence, linguistics and psychology 
are all integrated into complex procedures. 
The model I will describe in this paper considers 
the linguistic context as a means to restrict the 
large :number of interpretations that an expres- 
sion in principle has to a limited number of alter- 
natives. Knowledge of the world and/or prefer- 
ence of one alternative to the others may only re- 
strict the domain farther, never stretch it. This 
view will be defended in the paper. The former 
limitation is discussed under the name of 'possi- 
ble' analyses, the latter is christened 'preferred' 
analyses. 
I will first introduce the model in the next sec- 
tion. Sections 4 and 5 will motivate the approach 
and compare it to other proposals. Section 6 
will describe and discuss the so-called associa- 
tive anaphora and show that the model adopted 
can be extended to non-standard anaphoric ref- 
erence a.s well. 
3 The Model 
The model makes a distinction between possi- 
ble and preferred antecedents for a particu- 
lar a11aphor. The wellformedness definition be- 
low expresses the relation between an anaphor 
and its possible antecedents. Key notions in the 
definition are compatibility and recency. 
Definition 1 An antecedent Ant is a possible 
antecedent for an anaphor Ana iff 
- Ant and Ana are compatible 
- ~o antecedent compatible with Ana occurs iu a 
more recent unit 
The units referred to in the definition relate to 
units in the discourse. '\['he data structure re- 
95 
chap. 1 
Figure 1: The linguistic context 
flects the structure of the text, be it that only 
current units show internal structure. Pigure 1 
pictures the state of affairs at the time the 4th 
sentence in chapter 3, paragraph 3 of a book is 
being processed. The current chapter, paragraph 
and sentence show substructure, the others have 
no internal structure at all. 1 
Recency and compatibility are the central no- 
tions in the definition. The notion of recency 
needs no further explanation. Its importance 
for anaphora resolution has been demonstrated 
convincingly by ttobbs (1978). Compatibility 
demands agreement of linguistic features rele- 
vant to anaphora resolution. Pronoun resolution 
signals the need for features such as category, 
person, number, gender and humanness. Deft- 
nite NP resolution requires a more sophisticated 
mechanism. It should also be able to recognize 
relations such as synonymy, hyperonymy etc. 
Besides the wellformedness definition, one might 
want to adopt preference rules to select a (small 
set of) preferred antecedent(s) from the set 
of possible ones. I will not particularly be con- 
cerned with the preference rules in this paper, 
though I do foresee two areas that will deter- 
mine the nature of these rules. The first is world 
knowledge. In some cases, only extremely com- 
plex inferences can decide between two linguisti- 
cally equally preferred antecedents. The second 
is anaphoric preference. An anaphor sometimes 
prefers an object to a subject or a pronoun to a 
flfll NP, everything else being equal. The truth is 
1The model tmintendedly suggests that the units are 
syntactic in nature and in linear order. A more realis- 
tic picture includes semantic units and substructures, as 
argued for by a.o. Grosz(1977), Scha et.al.(1988) 
that everything else seldom is equal. Preference 
is very subtle and influenced by factors like sub- 
language, style etc. The discussion to follow will 
mainly be concerned with possible antecedents. 
4 Motivation 
The approach claims that anaphoric resolution 
is a two-stage procedure. The first stage de- 
termines the possible antecedents by means of 
tile linguistically defined notions of compatibil- 
ity and recency. The second stage consists of the 
calculation of preferred antecedents. I will first 
argue why the distinction between possible and 
preferred is important. And second, \[ will moti- 
vate why the definition of possible antecedents is 
expressed in terms of the notions compatibility 
and recency. 
The distinction possible vs. preferred is con- 
sidered essential for the following reasons. First, 
it is a means to determine whether a discourse 
is coherent or not, from the anaphoric point 
of view. Consider what happens when an an- 
tecedent is selected but needs to be rejected be- 
cause of information later in the discourse. As 
the meanings of terms like possible/impossible 
suggest, there is no way that we can backtrack 
on an initially possible antecedent in favour of an 
antecedent that is not part of the set of possible 
ones. Compare the following two discourses: 
(1) John was late for his appointment with Joe 
Fortunately, Joe was even later 
ttis work had kept him from leaving in time 
He, on the other hand, had missed the bus 
96 
(2) John was late for his appointment with 
Mary 
Fortunately, Mary was even later 
His work had kept him from leaving in time 
She, on the other hand, had missed the bus 
Discourse (1)is clearly incoherent whereas (2) 
is acceptable. The reasons are purely linguis- 
tic. The definition of possible antecedents above 
stated that some antecedents are not accessible 
to a pronoun because of other intermediate com- 
patible antecedents. Only the latter are possible 
candidates. When the continuation of the dis- 
course makes clear that they were not the right 
antecedents after all, the discourse should be 
considered incoherent. A preferred antecedent, 
however, may be rejected later on in the dis- 
course in favour of one that was not preferred, 
merely possible. An algorithm that merges pos- 
sible and preferred is not able to make this dis- 
tinction. 
The second motivation for maintaining the possi- 
ble/preferred distinction is provided by (linguis- 
tic) ambiguity in the language. Consider (3) and 
the examples cited in Winograd(1972), (4) vs. 
(3) The women met their husbands at a party 
They were very young at the time 
(4) The city councillors refused the demonstra- 
tors a permit because they feared violence 
(5) The city councillors refused the demonstra- 
tors a permit beca.use they advocated vio- 
lence 
Discourse (3) is three-way ambiguous and it is 
vague which of the meanings is intended. Blur- 
ring the distinction between preferred and possi- 
ble antecedents right away contradicts the intu- 
ition that all three antecedents in the first sen- 
tence might serve as antecedents for the pro- 
noun but not any NP that was mentioned be- 
fore this utterance. The examples (4) and (5) 
serve to show that linguistically possible is es- 
sentially different from possible as regards 'the 
world'. Assuming that no NLP-system, nor any 
human discourse participant by the way, has rich 
enough information to infer all essentials at the 
right time (yet), this approach at least yields a 
reliable, and finite, set of possible candidates. 
Having shown that the distinction between well- 
formed and preferred antecedent-anaphor rela- 
tions is necessary, it remains to be shown why 
possible antecedents are defined by means of the 
simple notions of compatibility and recency. The 
explanation is quite simple: they seern to yield 
the right results. It is well-known that full deft- 
nite NPs are in general further apart from their 
antecedents than pronouns are. Now consider 
that definite NPs, in general again, have more 
descriptive content than pronouns do. It fol- 
lows that pronouns may be compatible with an- 
tecedents that definite NPs do not match with. 
Together with the notion of recency, compati- 
bility then explains why the antecedents of defi- 
nite NPs may be at quite a distance. This only 
holds for definite NPs that have ample descrip- 
tive content though. An underspecified definite 
NP like 'the man' behaves similarly to the pro- 
noun 'he'. Both will accept the most recently 
mentioned male individual(s) as their possible 
antecedent(s). So, the number of units that the 
anaphor may search to find its antecedent is dy- 
namic rather than static. Consider discourse (1) 
versus (2) again. A static number of sentences 
would not explain the difference. The pronouns 
in (1) behave similar to those in (2) and yet, 
one of the two is incoherent. The reason is that 
'John' in (1) is not accessible because of the in- 
tervening NIP 'Joe'. The NP 'Mary' in (2) does 
not block reference to 'John' because 'Mary' is 
not compatible with the same anaphor. 
5 Comparison to Other Ap- 
proaches 
It might be argued that the notion of possible an- 
tecedents is hardly original in the literature on 
97 
the topic of anaphora resolution. The Focusing 
Approach reported on in Sidner (1983) also em- 
ploys a list that contains a number of antecedents 
from which the anaphor can select its co-referent. 
(cf. also Brennan et.al. 1987, Grosz 1983). Note, 
however, that the members on this list are sub- 
stantially different from the possible antecedents 
described in this paper. Potential antecedents 
in the Focusing Approach are not determined by 
the descriptive content of the anaphor but rather 
by what the discourse makes available for fur- 
ther reference. Suppose a sentence yields 'the 
pump' and 'the men' as available antecedents. 
They will be on the list from which a subsequent 
pronoun, say 'they', selects its antecedent. The 
difference is that 'the pump' may be potential 
in the Focusing Approach but in no way 'possi- 
ble' in the sense I used the term. Moreover, all 
potential antecedents in the Focusing Approach 
may be rejected in favour of an antecedent se- 
lected by a more global mechanism. This means 
that even the function of Sidner's list of potential 
antecedents in the discourse model differs from 
mine. I claim that failure in finding a co-referent 
on the list of possible antecedents corresponds to 
incoherence of the discourse. The list employed 
in the Focusing Approach has no such function. 
The anaphor could still find a referent on the 
more global part of the data structure, the so- 
called Focus Stack. 
6 Associative Anaphora 
To illustrate what the model has in store for a 
non-standard type of anaphoric reference, con- 
sider a case of the 'associative anaphoric use', as 
discussed in Hawkins(1978). A typical example 
of the phenomenon is in (6). 
(6) The man drove by in a car. The exhaust 
fumes were terrible. (cf. its exhaust fumes) 
Let us see how these constructions behave w.r.t. 
the model presented above. In view of the 
close correspondence between the bare associa- 
tive anaphor and an NP that includes the an- 
tecedent in pronominal form, I propose to anal- 
yse the constructions as having an empty pro- 
noun. The pronoun is anaphoric. Given the 
notion of compatibility and the fact that empty 
pronouns have no descriptive content at all, they 
are predicted to be compatible with any NP. 
Combining it with the notion of recency, it means 
that the set of possible antecedents consists of all 
and only the NPs in the most recent unit. To see 
whether this is correct, compare the following ex- 
amples: 
(7) I drove by our house in my car 
The windows were dirty 
(8) I drove by our house in my car 
I saw my father's car 
The windows were dirty 
(9) I drove by our house in my car 
I saw my father's bicycle 
The windows were dirty 
(10) I drove by our house in my car 
The windows were dirty 
The front door was open 
In (7), both 'our house' and 'my car' are pos- 
sible antecedents for the associative anaphor in 
the second sentence of the discourse. In (8) how- 
ever, only 'nay father's car' is possibly core i~r- 
ent with the empty pronoun. This indeed is 
what the model presented above predicts. Inter- 
vening potential antecedents block coreference 
with previous candidates. The NP 'my father's 
car' is compatible and in a more recent unit 
than 'our house' as well as 'my car'. Remem- 
ber that the antecedents we are talking about 
are possible antecedents in the sense of defini- 
tion 1 above. World knowledge or subsequent 
inibrmation in the discourse might still consider 
them inappropriate. This is illustrated in (9). It 
is 'my father's bicycle' that is the only possible 
antecedent. Ruling it out by world knowledge 
and having no possible alternative renders the 
discourse incoherent. 
4 
98 
Discourse (10) might seem to be a counterex= 
ample since 'the fi'ont door' can refer to 'the 
house' despite the occurence of 'the windows' in 
between. Ilowever, 'the windows' was not the 
only antecedent in the most recent unit. Re- 
member that we've assumed an empty pronoun 
all along. Among the most recent compatible an- 
tecedents is the empty pronoun that corefers to 
'house'. This explains why you can have chains 
of associative anaphora that superficially seem 
to corefer with an antecedent in a unit preced- 
ing the one in which the most recent compatible 
antecedent occurs. 
7 Conclusion 
The paper reported on a model that serves 
anaphora resolution. It might he characterized 
a~s a two-stage model. The first stage determined 
the set of linguistically possible antecedents. It 
was argned that compatibility and recency are 
adequate notions in determining the set of pos- 
sible antecedents. The task of the second stage 
is to restrict the set of possible antecedents to (a 
nurnber of) preferred antecedent(s). 
8 .Acknowledgments 
Louis des Tombe, thanks \[or your helpful sug- 
gestions and critical remarks. 1 was supported 
by the European Community and the NBBI 
through the Eurotra project. 
Conference on Artificial Intelligence, Cam-- 
bridge, Massachusetts, 1977. 

References 

\[1\] S. E. Brennan, M. W. l"riedman, and C. a. 
Pollard. A centering approach to pronouns. 
In Proceedings o f the 25th Annual Meeting of 
the Association for' Computational Linguis- 
tics, 1987. 

\[2\] 13. Grosz. The Representation and Use of Pc> 
cus in a System for Understanding Dialogues. 
In Proceedings of the fifth International Joint

\[3\] J. tlawkins. Definiteness and Indefiniteness. 
London: Croom IIelm, 1978. 

\[4\] I. Iteim. The Semantics of Definite and in- 
definite Noun Phrases. Unpublished PhD 
dissertation, University of Massachusetts, 
Amherst, 1982. 

\[5\] a. Ilobbs. Resolving pronoun references. Lin- 
gua, 44, 1978. 

\[6\] II. Kamp A theory of truth and semantic 
interpretation. In J. Groenendijk et.al. (eds.) 
Formal Methods in the Study of Language. 
Amsterdam: Mathematical Centre, 1981. 

\[7\] R. Scha and L. Polanyi. An Augmented Con- 
tex Free Grammar for Discourse. In Proceed- 
ings of Coling Budapest, 1988. 

\[8\] C. L. Sidner. Focusing in the comprehen- 
sion of definite anaphora. In M. Brady and 
R. C. Berwiek (eds.), Computational Models 
of Discourse. MIT press, 1983. 

\[9\] T. Winograd. Understanding Natural Lan- 
guage. New York: Academic Press, 1972. 
