Proceedings of EACL '99 
Pointing to Events 
Frank Schilder 
Department for Informatics 
University of Hamburg 
Vogt-KSlln-Str. 30 
22527 Hamburg 
GERMANY 
schi!der@informatik, uni- hamburg, de 
1 Introduction 
Although there is an extensive body of research 
concerned with anaphora resolution (e.g. (Fox, 
1987; Grosz et al., 1995)), event anaphora has 
been widely neglected. This paper describes 
the results of an empirical study regarding event 
reference. The experiment investigated event 
anaphora in narrative discourse via a sentence 
completion task. The results of the experiment 
suggest that the discourse structure established by 
an explanation or elaboration relation has an in- 
fluence on whether the last mentioned event, pre- 
viously mentioned events or a complex event is 
preferably referred to. 
First, a short introduction is given to the ob- 
servation that reference by the demonstrative pro- 
noun this can only be done to antecedents men- 
tioned in segments at the right frontier of the dis- 
course structure tree (Webber, 1991). Second, al- 
though the right frontier restriction is generally 
agreed on by researchers, it is still unclear what 
kind of information is actually stored on higher 
levels of the discourse tree. Segmented Discourse 
Structure Theory (SDRT) is introduced (Asher, 
1993) and the predictions of this theory are dis- 
cussed regarding event anaphora for two example 
discourses. These predictions are also compared 
with the predictions according to Webber (1991). 
Neither theory adequately explains the results 
of the current experiment. Thus the presentation 
concludes with an outline of a proposed computa- 
tional theory. 
2 Background 
Generally speaking, events are referred to by 
pronominal expressions such as that or this: 1 
(1) (a) John hated snakes. (b) His neighbour 
had kept snakes, (c) and he had been bit- 
1The sequence is a slightly altered example taken 
from Webber (1991). 
ten by a snake once. (d) This was unfor- 
tunate. 
The pronoun this in (ld) may refer to the en- 
tire situation described by (la) through (lc). But 
there are other conceivable antecedents for this. 
The situation described by (lc) may be referred 
to by this as well, if we consider an alternation of 
(ld) as in the following: 
(1') d'. This was quite dangerous. 
The example discourse illustrates that the dis- 
course in (la-c) is organised hierarchically and 
consists of different discourse segments (e.g. (la- 
c) and (lb-c)). Webber (1991) points out that 
events are only available for anaphoric reference 
when they are mentioned by the last utterance 
(e.g. (lc)) or by the situation that is constructed 
by the preceding discourse segment (e.g. (1a-c)). 
The right frontier hypothesis reflects this observa- 
tion as shown by bold typeset in figure 1. The dis- 
(1 (a,(b,c))) 
(la) (l(b,c)) 
(lb) (lc) 
Figure 1: The discourse structure for (1a-c) 
course structure also influences how we conceptu- 
alise situations described in a text. In particular, 
a segment in (la-c) establishes a complex event 
that can be referred to. 
3 Predictions 
Unfortunately, there is much disagreement be- 
tween theoretical approaches to discourse process- 
ing on what exactly is available on the right fron- 
tier. The tree structure proposed by Webber, for 
example, contains the semantic content grouped 
together for the entire segment. 
283 
Proceedings of EACL '99 
'This refers to explanation elaboration 
segment: (e.g. (2)) (e.g. (3)) 
a 
b 
C 
(b-c) 
(a-c) 
2.08% 
9.72% 
69.44% 
18.06% 
0.69% 
9.49% 
37.97% 
38.61% 
13.29% 
0.63% 
Table 1: Results of the sentence completion task 
Asher (1993), on the other hand, defines open 
attachment sites of the discourse structure by the 
term openness via rhetorical relations. Only elab- 
oration and explanation allow reference to an- 
tecedents mentioned in sentences other than the 
current one. An explanation relation holds, for 
example, between (2b) and (2c) (and elaboration 
between (3b) and (3c) see below): 
(2) (a) Peter was sitting in his chair at home. 
(b) All of a sudden, he stood up. (c) The 
doorbell had rung. This was... 
Following Webber's account, this can only refer to 
the situation described by the last utterance (2c) 
or the situations described by the sentences (2b-c) 
or (2a-c), but not to (2a) and (2c). 
In contrast, SDRT predicts that the situations 
described by sentence (2b) or (2c) are available, 
but not by (2a) nor any combined situation. 
Consequently, the empirical investigation fo- 
cuses especially on the question of which described 
events are available for further attachments and 
event anaphora depending on the chosen rhetori- 
cal relation (i.e. explanation or elaboration). 
4 Experimental data 
Six ambiguous sequences such as (2) or (3) were 
presented to 115 participants. They were asked 
first to continue the sentences and then to specify 
explicitly what this referred to. 
(3) Saturday was Jerry's birthday. Several 
students organized a dinner for Jerry. 
Some students had written fancy invita- 
tion cards. This was... 
Note that the results as shown by table 1 can- 
not be explained by either theoretical approach. 
The results for the explanation relation is best ex- 
plained by Webber, since either the last mentioned 
situation or a combined situation of the segments 
(2b-c) were most often chosen by the participants. 
On the other hand, results for the elaboration re- 
lation contradict the predictions made. The sit- 
uation described by (3b) should not be available 
according to Webber's account, instead an ambi- 
guity between (3b) and (3c) can be observed for 
this sequence. This would be predicted by SDRT. 
But SDRT cannot explain the combination of two 
situations that was chosen by subjects 18% of the 
time for an explanation relation. 2 
Results suggest that the type of rhetorical re- 
lation has a crucial influence. An explanation re- 
lation linking (2b) and (2c) focuses on the sen- 
tence describing the cause (i.e. the last utterance), 
whereas an elaboration relation as in (3) estab- 
lishes an ambiguous sequence between the last two 
described situations. In addition, note that a sig- 
nificant proportion of the participants referred to 
a combined event regardless of the rhetorical rela- 
tion. 
5 A computational model 
A computational model could benefit from the re- 
sults presented in this paper. First, the choice of 
rhetorical relation should affect the way possible 
antecedents are represented for anaphora resolu- 
tion. Second, the current data as well as data from 
reading experiments, for instance, could be used 
to determine the likelihood of possible attachment 
sites in the discourse structure. 
However, another question remains: what 
should the representation of a complex event look 
like? The tree structure proposed by Webber con- 
tains the semantic content grouped together for 
the entire segment. It seems more plausible that 
at a higher level of a discourse some form of ab- 
straction takes place. 
Future research must focus on the information 
that is available from preceding text and what 
is obtainable for anaphoric expressions such as 
demonstrative pronouns and definite descriptions. 
References 
Nicholas Asher. 1993. Reference to abstract Ob- 
jects in Discourse. Kluwer Academic Publish- 
ers, Dordrecht. 
Barbara Fox. 1987. Discourse structure and 
anaphora. Cambridge University Press. 
Barbara J. Grosz, Aravind Joshi, and Scott Wein- 
stein. 1995. Centering: A framework for mod- 
elling the local coherence of discourse. Compu- 
tational Linguistics, 21(2):203-225. 
Bonnie L. Webber. 1991. Structure and ostension 
in the interpretation of discourse deixis. Lan- 
guage and Cognitive Processes, 6(2):107-135. 
2Stories within each category differed in the num- 
ber of subjects referring to a combined situation (b-c). 
The numbers presented here are the averages across 
the stories. 
284 
