Long Distance Pronominalisation and Global Focus 
Janet Hitzeman and Massimo Poesio 
CSTR and HCRC, University of Edinburgh 
2 Buccleuch Place, Edinburgh EH8 9LW, Scotland 
{J. Hi t z eman, Mass imo. Poes io} @ed. ac. uk 
Abstract (1) 
Our corpus of descriptive text contains a signifi- 
cant number of long-distance pronominal references 
(8.4% of the total). In order to account for how these 
pronouns are interpreted, we re-examine Grosz and 
Sidner's theory of the attentional state, and in par- 
ticular the use of the global focus to supplement 
centering theory. Our corpus evidence concerning 
these long-distance pronominal references, as well 
as studies of the use of descriptions, proper names 
and ambiguous uses of pronouns, lead us to con- 
clude that a discourse focus stack mechanism of the 
type proposed by Sidner is essential to account for 
the use of these referring expressions. We suggest 
revising the Grosz & Sidner framework by allowing 
for the possibility that an entity in a focus space may 
have special status. 
1 Motivation 
We call LONG-DISTANCE PRONOMINALISATIONS 
those cases of pronoun use in which the antecedent 
of the pronoun does not occur either in the same 
sentence as the pronoun or in the immediately pre- 
ceding one, but further back in the text. These cases 
are thought to be rare on the basis of studies such 
as (Hobbs, 1978), which found that 98% of pro- 
noun antecedents in the corpus analysed were in the 
same sentence as the pronoun or the previous one. 
However, our analysis of a small corpus of oral de- 
scriptions of museum items collected for the ILEX 
project (Hitzeman et al., 1997) revealed that long- 
distance pronouns are much more common in this 
kind of data -four times as common, in fact: out of 
a total of 83 pronouns, 7 (8.4%) were long-distance. 
The typical pattern of long-distance pronominalisa- 
tion in the ILEX dialogues is shown in (1), where the 
pronoun him in the last sentence refers to the jew- 
eller, mentioned most recently two sentences ear- 
lier. 
JO: Okay, thank you. Shall we look at the object in 
case number 16, number 1 ? There's a set of three 
objects here. 
LG: 1. Yes. 
2. What these symbolise for me are one of the pre- 
occupations of the 1980s, which is recycling. 
3. The jeweller who made these bangles was par- 
ticularly interested in the idea of using intrinsically 
worthless material- material that had been thrown 
away, old junk- and he lavished on those materials 
an incredibly painstaking and time-consuming tech- 
nique, so that the amount of time put into the labour 
of making these jewels bears absolutely no relation 
to the value of the materials that he's used. 
4. And if you look at, for instance, the bangle at the 
bottom- that's the blue and red one- what looks as 
though it's painted decoration is in fact inlaid; it's 
bits of cut-off razor-blade, biro, knitting needles, in- 
laid into layer after layer of resin, which is done in 
emulation of Japanese lacquer technique. 
5. And that particular bangle took hhn something 
like 120 hours of work. 
All 7 long-distance pronouns in the ILEX dialogues 
we have studied refer to discourse entities intro- 
duced in background text in this way. 
Unlike Sidner's theory of focus (Sidner, 1979), 
the theory of the attentional state in (Grosz and 
Sidner, 1986) (henceforth: G&S) does not include 
explicit provision for long-distance pronominalisa- 
tions, although some of the necessary tools are po- 
tentially already there, as we will see. The compo- 
nent of the theory that deals with pronominal ref- 
erence, centering theory (Grosz et al., 1995), only 
accounts for cases in which the antecedent of a pro- 
noun is introduced by the previous sentence; cases 
such as (1) have to be handled by different mech- 
anisms. In this paper we look the phenomenon of 
long-distance pronominalisation in some detail, ex- 
amining data from different domains, and consider 
550 
its implications for G&S's theory. 
2 Theories of focus 
Space unfortunately prevents a full discussion of 
Grosz's (1977), Sidner's (1979), and G&S's (1986) 
theories of focus and the attentional state in this 
abstract. The crucial aspects of these theories, for 
the purpose of the discussion below, are as follows. 
First of all, G&S propose a distinction between two 
components of the attentional state: the GLOBAL 
FOCUS, structured as a stack of focus spaces and ac- 
cessed to interpret definite descriptions; and the LO- 
CAL FOCUS, consisting of the information preferen- 
tially used to interpret pronouns• In addition, they 
adopt CENTERING THEORY (Grosz et al., 1995) as 
a theory of the local focus. 
Secondly, although G&S's theory integrates 
ideas from both Grosz's and Sidner's original the- 
ories, and although both of these theories assumed 
a stack structure, the global focus in G&S's the- 
ory is structured as a stack of focus spaces, as in 
Grosz's original proposal, rather than as a stack of 
discourse foci, as in Sidner's original theory. The 
claim that different parts of the attentional state are 
accessed when resolving pronouns and definite de- 
scriptions is supported, broadly speaking, by psy- 
cholinguistic research (see, e.g., (Garrod, 1993)). 
The main claims of centering theory are also con- 
sistent with psycholinguistic results (Hudson, 1988; 
Gordon et al., 1993). To our knowledge, however, 
the choice of a stack of focus spaces over a stack 
of discourse foci has never been motivated; yet this 
decision plays a crucial role in our problem, as we 
will see. 
A point worth keeping in mind throughout the 
following discussion is that, although the concept 
• O" of c8 (centerm~ theory s name for the current most 
salient entity) was originally introduced as 'roughly 
corresponding to Sidner's concept of discourse fo- 
cus', in fact it is not clear that the two concepts are 
capturing the same intuitions (Poesio and Steven- 
son, 1998). Although it is often the case that the CB 
and the discourse focus coincide, this is not true in 
general.I For the purposes of this paper, however, 
we will assume that the two notions do coincide, 
and will use the neutral term MOST SALIENT EN- 
TITY (MSE) to refer to the discourse focus / CB of a 
particular segment of text. 
~This intuitive impression was confirmed by a recent study 
(Giouli• 1996), whose author tracked both the 'intuitive CB' and 
the 'intuitive discourse focus' of 8 Map Task conversations. 
3 The Data 
The Intelligent Labelling Explorer (ILEX) project is 
building a system that generates descriptions of ob- 
jects displayed in a museum gallery. 2 In order to 
generate the most natural descriptions of the ob- 
jects, dialogues with a museum curator were col- 
lected, describing objects in the National Museum 
of Scotland's 20th Century Jeweilery Gallery. We 
will refer to this corpus as the ILEX corpus. In the 
dialogues, the curator (LG) moves from case to case 
as directed by an observer (JO) and describes the 
jewels in each case, as in example (1). 
The work described here is part of two related 
projects: SOLE, the goal of which is to extend 
the ILEX system with the capability of generating 
prosodically adequate speech, and GNOME, which 
is concerned with the generation of nominal expres- 
sions. A second corpus of museum object descrip- 
tions was collected for use with SOLE; we will refer 
to this corpus as the SOLE COrpUS. 
4 Analysis 
4.1 First Hypothesis 
Because G&S's theory of the attentional state al- 
ready hypothesises global focus structures in addi- 
tion to the local attentional structures assumed in 
centering theory, the simplest explanation for our 
cases of long-distance pronominalisation is to hy- 
pothesise that readers exploit the global focus to re- 
solve pronouns in such cases. 
Assuming that the global focus is involved in 
these cases, instead of complicating the local fo- 
cus/centering theory, is consistent with the little 
available psychological evidence --e.g., with the re- 
suits of Clark and Sengui (1979), who observed a 
slowdown in reading times for the sentence con- 
taining the pronoun when the antecedent is not in 
the same or the previous sentence, implying that 
long-distance pronominal anaphora are handled dif- 
ferently. 
Furthermore, suggesting that these pronouns are 
resolved by accessing the global focus would not 
really amount to a revision of the basic assump- 
tions of G&S's theory. Although no explicit pro- 
posal conceming the respective roles of local fo- 
cus and global focus in pronoun resolution has ever 
been made in the literature on the G&S framework, 
cases of pronouns involving access to the global at- 
tentionai structure rather than to the local focus have 
~'http: //www.cogsci .ed.ac.ukfalik/ilex/ 
systemintro, html 
551 
already been discussed in this literature. So-called 
RETURN-POPS, which are pronouns that signal a re- 
turn to a superordinate discourse segment, were dis- 
cussed by Grosz (1977) and then in detail in (Fox, 
1987). In (2), for example, sentence 5 resumes the 
segment interrupted by 2-4; the antecedent for the 
pronoun her is supposed to be found on the stack, 
although the details of this process have never re- 
ally been discussed. 3 
(2) 1. C: Ok Harry. I'm have a problem that 
uh my-with today's economy my daughter is 
working, 
2. H: I missed your name. 
3. C: Hank. 
4. H: Go ahead Hank. 
5. C: as well as he___zr husband 
A second case of pronouns that clearly seem to in- 
volve access to some global structure are so-called 
'generic' pronouns, such as the, in (3): 
(3) John went back to the hotel. He looked for 
Mary in their room, but couldn't find her. They 
told him that she had left about an hour earlier. 
(We are not aware of any account of these uses of 
pronouns within the G&S framework.) 
As we will see in a moment, the long-distance 
pronouns observed in the ILEX dialogues are neither 
generic pronouns nor return-pops; nevertheless, we 
are going to show that these cases, as well, are re- 
solved by accessing the global focus. 
4.2 Long-distance pronouns need not be 
return-pops 
Tile use of him in the last sentence of (1) could only 
be termed a RETURN-POP if it were to involve a re- 
turn to the previous discourse segment which 'pops 
over' sentence 4 (And if you look at, for instance, the 
bangle at the bottom ... ) and 'closes off' the ma- 
terial introduced in that sentence. But this is clearly 
not the case, as shown by the fact that the final sen- 
tence contains a reference to both the jeweller and 
the bangle. Indeed, the bangle could also be referred 
to with a pronoun: And it took him something like 
120 hours of work. The fact that pronouns and def- 
inite NPs in the last sentence can refer back to ma- 
terial in the 4th sentence indicates that this material 
must still be on the stack. 
3This example is from (Pollack et al., 1982). 
4.3 Discourse Structure in the Example Text 
Before discussing how the global focus is used for 
resolving pronouns such as the long-distance pro- 
noun in the last sentence of (1), we need to discuss 
the structure of these examples: i.e., is the part of 
(l) which has the jeweller as MSE (2nd sentence) 
still on the stack when the part that describes details 
of the jewel and contains the long-distance pronoun 
(3rd and 4th sentence) is processed? 
Answering this question is made more difficult by 
the fact that G&S's theory of the intentional struc- 
ture is very abstract, and therefore does not help 
much in specific cases, especially when the genre 
is not task-oriented conversations. More specific 
indications concerning the structure of the relevant 
example, and more in general of the conversations 
in the ILEX corpus, are given by Rhetorical Struc- 
ture Theory (RST) (Mann and Thompson, 1988), 4 
although even with RST it is still possible to anal- 
yse any given text in many different ways. Nev- 
ertheless, we believe that the structure depicted in 
Figure 1 is a plausible analysis for (1); an alterna- 
tive analysis would be to take the 4th and 5th sen- 
tence as elaborations of and he lavished on those 
materials an incredibly painstaking technique .... 
but in this case, as well (and in all other rhetorical 
structures we could consider) sentences 4 and 5 are 
satellites of sentence 3. (We have employed the set 
of rhetorical relations currently used to analyse the 
ILEX data.) 
The relation between G&S's and RST's notion 
of structure has been analysed by, among others, 
(Moore and Paris, i 993; Moser and Moore, 1996). 
According to Moser and Moore, the relation can be 
characterised as follows: an RST nucleus expresses 
an intention I~; a satellite expresses an intention 18; 
and I,~ dominates Is. Thus, in (1), the nucleus of the 
exemplification relation, sentence 3, would domi- 
nate the satellite, consisting of sentences 4 and 5. 
We will make the same assumption here. Hence 
we can assume that the third sentence in (1) will 
still be on the stack when processing the 4th and 5th 
sentences. This would also hold for the alternative 
rhetorical structures we have considered. 5 
4Fox, as well, used RST to analyse the structure of texts in 
her study of the effect of discourse structure on anaphora (Fox, 
1987). 
5Some readers might wonder whether it wouldn't be sim- 
pler to assume that all of the utterances in (1) are part of the 
same segment. This assumption would indeed make the an- 
tecedent accessible; however, it would not explain the data. not 
at least if we assume that it is centering theory that determines 
552 
'jeweller particularly 
interested in using 
worthless material' 
The jeweller who made these 
bangles .... 
EXEMPLIFICATION 
And if you look at, for instance .... 
And that particular bangle took HIM .... 
CONSEQUENCE 
so that the amount of time he 
put into the labour .... 
Figure 1: A possible analysis of (1). 
4.4 What Goes on the Stack? 
We can finally turn to the task of explaining how 
the global focus is used to resolve long-distance 
pronominalisations. The simplest explanation con- 
sistent with the G&S's framework would be to as- 
sume that resolving such pronouns involves search- 
ing for the first discourse entity in the focus space 
stack that satisfies gender and number constraints. 
Under the assumptions about the discourse struc- 
ture of examples like (1) just discussed, this expla- 
nation would indeed account for that example; there 
is evidence, however, that additional constraints are 
involved. The first bit of evidence is that the pres- 
ence on the focus space stack of an appropriate an- 
tecedent does not always make the use of a long 
distance pronoun felicitous. Consider tile follow- 
ing fi'agment of an article that appeared in The 
Guardian, January 28, 1995, p.3. 
(4) Joan Partington, aged 44, from Bolton, Lan- 
cashire, has six children. The eldest are two 
17-year-old twin boys, one awaiting a heart by- 
pass operation and the other with severe be- 
havioral problems. A 13-year-old son has hy- 
drocephalus. She was living with her hus- 
band when Wigan magistrates ordered her to 
be jailed unless she paid £5 per week, although 
he earned only £70 per week as a part-time 
postman. 
anaphoric reference, since centering does not explain how a 
pronoun can refer to an antecedent two sentences back. Assum- 
ing that there is more than one segment in such texts, instead, 
will turn out to be not just a more plausible assumption about 
segmentation; it will also give us a simple way to explain the 
data, 
The use of he in the last sentence is awkward, even 
though there is a discourse entity on the focus space 
stack- the husband- that would satisfy the con- 
straints imposed by the pronoun. This seems to in- 
dicate that the elements of a focus space are not all 
equally accessible. 
The second relevant bit of evidence concerns the 
use of proper names in the ILEX corpus. It may hap- 
pen in the ILEX dialogues that a designer like Jessie 
King is first mentioned by name in a segment where 
she is not the main topic of discussion, as in Other 
jewels in the Bohemian style include a brooch by 
Jessie King. If this is the case, then when later we're 
talking about another jewel that King designed, she 
will have to be introduced again with a full proper 
name, Jessie King, rather than simply King. If, how- 
ever, she becomes the 'main topic' of discussion, 
then later, whenever we talk about her again, we can 
use reduced forms of her proper name, such as King. 
Again, this difference is not easy to explain in terms 
of focus spaces if we assume that all objects in a 
focus space have the same status. 
A third class of expressions providing evidence 
relevant to this discussion are bridging descriptions, 
i.e., definite descriptions like the door that refer to 
an object associated with a previously mentioned 
discourse entity such as the house, rather than to 
the entity itself (Clark, 1977). Poesio et al. (1997; 
1998) report experiments in which different types of 
lexical knowledge sources are used to resolve bridg- 
ing descriptions and other cases of definite descrip- 
tions that require more than simple string match for 
their resolution. Their results indicate that to re- 
solve bridging descriptions it is not sufficient sim- 
ply to find which of the entities in the current focus 
553 
space is semantically closest to the bridging descrip- 
tion: in about half of the cases of bridging descrip- 
tions that could be resolved on the basis of the lexi- 
cal knowledge used in these experiments, the focus 
spaces contained an entity whose description was 
more closely related to that of the bridging descrip- 
tion than the one of the actual antecedent(s). 
This evidence about infelicitous pronouns, proper 
names, and bridging descriptions suggests that the 
entities in a focus space are not all equally salient. 
In fact, one could even wonder if we need focus 
spaces at all; i.e., if Sidner's original proposal - ac- 
cording to which it's just the MSE that goes on the 
stack, not the whole focus space - is correct. A re- 
vision of G&S's theory along these lines- i.e., in 
which the focus space stack is replaced by an MSE 
stack- would still explain (1), since the jeweller is 
clearly the MSE of sentence 3; indeed, all 7 cases 
of long-distance pronouns found in the ILEX corpus 
have a previous MSE as their antecedent. But, in ad- 
dition, this revision would explain the awkwardness 
of (4): the husband was never an MSE, so it would 
not be on the stack. 
A global focus of this type would also give us a 
way to formulate a restriction on using shortened 
forms of proper names that would account for the 
facts observed in the ILEX corpus: reduced NPs 
are allowed for entities that have been introduced 
as MSEs, full NPs are needed otherwise. And fi- 
nally, keeping track of previous MSEs seems essen- 
tial for bridging descriptions as well: in order to 
find the reasons for the low performance of algo- 
rithms for resolving bridging descriptions entirely 
based on lexical knowledge, (Poesio et al., 1998) 
examined the bridging descriptions their corpus to 
find out their 'preferred' antecedent. 6 They found 
that the preferred antecedent of a bridging descrip- 
tion is a previous MSE in 54 out of 203 cases. In 
the SOLE COrpUS, 8 OUt of 11 bridging descriptions 
relate to the MSE. 
Does this mean, then, that we can get rid of fo- 
cus spaces, and assume that it's MSEs that go on the 
stack? Before looking at the data, we have to be 
clear as to what would count as evidence one way 
or the other. Even an approach in which only previ- 
ous MSES are on the stack would still allow access to 
entities which are part of what Grosz called the IM- 
PLICIT FOCUS of these MSEs, i.e., the entities that 
6As discussed in (Poesio and Vieira, 1998), in general there 
is more than one potential 'antecedent' for a bridging descrip- 
tion in a text. 
are 'strongly associated' with the MSES. This notion 
of 'strong association' is difficult to define- in fact, 
it is likely to be a matter of degree- but nevertheless 
it is plausible to assume that the objects 'strongly 
associated' with a discourse entity A do not include 
every discourse entity B which is part of a situation 
described in the text in which A is also involved; and 
this can be tested with linguistic examples, up to a 
point. For example, whereas definite descriptions 
like the radiator cap can easily be resolved in a null 
context to a car, descriptions like the dog can't, as 
shown by the infelicity of (5d) as a continuation of 
(5b), even though dogs in cars are not uncommon; 
some contextual antecedent is needed. 
(5) a. Mary saw a dark car go by quickly. 
b. It was a bright, warm day. 
c. The radiator cap was shining in the sun. 
d. The dog was enjoying the warmth. 
The question we have to answer, then, is whether 
the only information that is available as part of the 
attentional state is what is 'strongly associated' with 
one of the previous MSES, or, instead, all of the in- 
formation mentioned in the text. 7 
Now, sentences like (5a) license both bridging de- 
scriptions to the car, as in (5c), and to Mary, as in 
Her hat had become very hot. Whatever we take 
the MSE of (5a) tO be, it seems implausible to ar- 
gue that both the bridging description s the radiator 
cap and Her hat are resolved by looking at the ob- 
jects 'strongly associated' with that discourse entity. 
It is much simpler to assume that both Mary and 
the car are still accessible as part of the focus space 
constructed to represent the situation described by 
the text. This also holds for what we have called 
'generic' pronouns, as shown by (3), in which they 
refers to individuals associated with the hotel men- 
tioned in the first sentence, not to the MSE, John. 
And indeed, Sidner assumed two stacks- one of 
discourse foci, the other of actor foci. But even 
this extension would not be enough, because the an- 
tecedent of a bridging description is not always an 
entity explicitly introduced in the text, but can also 
be a more abstract DISCOURSE TOPIC, by which we 
7Notice however that the claim that only MSES go on the 
stack does not entail that everything else in the text is simply 
forgotten- the claim is simply that that intbrmation is not avail- 
able for resolving references anymore; presumably it would 
be stored somewhere in 'long term memory'. Conversely, 
the claim that everything stays on the stack would have to be 
supplemented by some story concerning how information gets 
forgotten-e.g., by some caching mechanism such as the one 
proposed by Walker (1996). 
554 
mean an issue / proposition that can be said to char- 
acterise the content of the focus space as a whole. In 
a corpus analysis done in connection with (Poesio et 
al., 1997; Poesio et al., 1998), we found that 7 out 
of 70 inferential descriptions were of this type; in 
the SOLE corpus, in which 3 out of 11 bridging de- 
scriptions behave this way. An example of this use 
is the description the problem below, that refers to 
the problem introduced by the first sentence in the 
text: 
(6) Solo woodwind players have to be creative if 
they want to work a lot, because their reper- 
toire and audience appeal are limited .... The 
oboist Heinz Holliger has taken a hard line 
about the problem ... 
Reference to abstract objects in general seem to re- 
quire maintaining information about the events and 
situations described by a text on the stack- see, e.g., 
(Webber, 1991). So, it looks like what we need is 
something of a compromise between the notion of 
global focus implicit in Sidner's original proposal 
and that proposed by G&S. 
4.5 The proposal 
The following hypothesis about the global focus and 
its use in pronoun resolution seems to provide the 
best account of the evidence we have examined: 
1. The global focus consists of a stack of fo- 
cus spaces, as in G&S's proposal. Each of 
these focus spaces can be summarised as be- 
ing 'about' some object / proposition / issue- 
indeed, more than one- for which we will use 
the term DISCOURSE TOPICS; but, in addition, 
2. Each focus space may be optionally associated 
with a MOST SALIENT ENTITY (MSE) explic- 
itly introduced in the text. 
3. The antecedent for a non-generic pronoun is 
preferentially to be found in the local focus; if 
none is available, one of the MSEs associated 
with a focus space on the stack can also serve 
as antecedent. 8 
4. Generic pronouns refer back to the situation 
described by the current focus space; 
5. Bridging descriptions can be related either to 
an entity in the current focus space, or to an 
MSE, or to a discourse topic; 
rThis would explain the difference in reading times ob- 
served by (Clark and Sengul, 1979). 
6. Definite descriptions can refer back to any en- 
tity in the global focus, including discourse 
topics. 
The reason for using the term 'optional' in 2 is that 
whereas focus spaces can always be described as be- 
ing about something, they are not always associated 
with a 'most salient entity': e.g., the first sentence 
in (6) introduces several topics (woodwind players, 
their need to be creative, etc.) but does not introduce 
an MSE. 
5 Related Work 
In a recent paper, Hahn and Strube (1997) propose 
to extend centering theory with what is, essentially, 
Sidner's stack of discourse foci, although their al- 
gorithm for identifying the ce is not identical to 
Sidner's. Their analysis of German texts shows a 
rather good performance for their algorithm, but, as 
only MSEs are predicted to be accessible, none of 
the anaphors depending on focus space information 
could be resolved. Their algorithm also appears to 
treat definite descriptions and pronouns uniformly 
as 'anaphors', which seems problematic in the light 
of psychological evidence showing that they behave 
differently, and examples like the following: 
(7) a. John/saw Mary. He/greeted her. 
b. John/saw Mary. ??The mani greeted her. 
(Guindon, 1985) proposed an alternative model 
of the attentional state involving a cache instead of 
a stack, and Walker (1996) argues that the cache 
model can account for all of the data that origi- 
nally motivated the stack model and, in addition, 
explains the use of informationally redundant ut- 
terances. The cache model isn't yet specified in 
enough detail for all of its implications for the data 
discussed here to be clear, but it appears that some 
of the issues discussed in this paper would have to 
be addressed in a cache model as well, and that 
some of our conclusions would apply in a model 
of that type as well. In particular, these propos- 
als are not very specific about whether the cache 
should count as a replacement of just the global fo- 
cus component of G&S's theory or of both local 
and global focus, and about what should go in the 
cache-Guindon seems to assume that it's discourse 
entities, whereas Walker also seems to allow for 
propositions and relational information. If the cache 
was intended as an alternative model of the global 
focus component, the data discussed here could be 
555 
taken as an argument that what goes in the cache 
should be focus spaces with distinguished MSEs. 
6 Conclusions 
Our main intent in looking at long-distance pronom- 
inalisation was to make some of the aspects of the 
G&S model of attentional state more precise, and 
to clarify its connection with earlier work by Sid- 
ner. The evidence we have presented suggests a 
main conclusion and a corollary. The main conclu- 
sion is that the uses of long-distance pronouns in our 
corpus can be explained as cases of reference to the 
MSE of a segment whose associated focus space is 
still on the stack. The corollary is that these ex- 
amples can be accounted for within a G&S-style 
model of discourse structure, provided that the the- 
ory is augmented by singling out some entities in 
focus spaces, and having these entities do some of 
the work done by Sidner's stack of discourse foci. 
A concern with studies of this type is that notions 
such as 'most salient entity' are hard to define, and 
it's not obvious that two different researchers would 
necessarily agree on what is the MSE of a given sen- 
tence. Work on verifying whether the notion we are 
assuming can indeed be reliably identified is under 
way as part of the GNOME project. 

References 
H. H. Clark and C. J. Sengul. 1979. In search of refer- 
ents for nouns and pronouns. Memory and Cognition, 
7(I ):35-4 1. 
H. H. Clark. 1977. Bridging. In P. N. Johnson-Laird and 
P.C. Wason, editors, Thinking: Readings in Cognitive 
Science. Cambridge University Press. 
B. A. Fox. 1987. Discourse Structure and Anaphora. 
Cambridge University Press, Cambridge, UK. 
S. Garrod. 1993. Resolving pronouns and other 
anaphoric devices: The case for diversity in dis- 
course processing. In C. Clifton, L. Frazier, and 
K. Rayner, editors, Perspectives in Sentence Process- 
ing. Lawrence Erlbaum. 
P. Giouli. 1996. Topic chaining and discourse structure 
in task-oriented dialogues. Master's thesis, University 
of Edinburgh, Linguistics Department. 
P. C. Gordon, B. J. Grosz, and L. A. Gillion. 1993. Pro- 
nouns, names, and the centering of attention in dis- 
course. Cognitive Science, 17:311-348. 
B. J. Grosz and C. L. Sidner. 1986. Attention, inten- 
tion, and the structure of discourse. Computational 
Linguistics, 12(3): 175-204. 
B. J. Grosz, A. K. Joshi, and S. Weinstein. 1995. Center- 
ing: A framework for modelling the local coherence 
of discourse. ComputationalLinguistics, 21(2):202- 
225. 
B. J. Grosz. 1977. The Representation and Use of Fo- 
cus in Dialogue Understanding. Ph.D. thesis, Stan- 
ford University. 
R. Guindon. 1985. Anaphora resolution: Short-term 
memory and focusing. In Proc. of the 23rd Annual 
Meeting of the ACL, pp. 218-227. 
U. Hahn and M. Strube. 1997. Centering in-the-large: 
Computing referential discourse segments. In Proc. 
of the 35th Annual Meeting of the ACL, pp. 104-111. 
J. Hitzeman, C. Mellish, and J. Oberlander. 1997. Dy- 
namic generation of museum web pages: The intelli- 
gent labelling explorer. Archives and Museum lnfor- 
matics, 11:107-115. 
J. R. Hobbs. 1978. Resolving pronoun references. Lin- 
gua, 44:311-338. 
S.B. Hudson. 1988. The Structure of Discourse and 
Anaphor Resolution: The Discourse Center attd the 
Roles of Nouns and Pronouns. Ph.D. thesis, Univer- 
sity of Rochester. 
W. C. Mann and S. A. Thompson. 1988. Rhetorical 
structure theory: Towards a functional theory of text 
organization. Text, 8(3):243-281. 
J. D. Moore and C. L. Paris. 1993. Planning text for ad- 
visory dialogues: Capturing intentional and rhetorical 
information. Computational Linguistics, 19(4):651- 
694, December. 
M. Moser and J. Moore. 1996. Toward a synthesis of 
two accounts of discourse structure. Contputational 
Linguistics, 22(3):409-419. 
M. Poesio and R. Stevenson. 1998. Computational mod- 
els of salience and psychological evidence. In prepa- 
ration. 
M. Poesio and R. Vieira. 1998. A corpus-based investi- 
gation of definite description use. ComputationalLin- 
guistics. To appear. 
M. Poesio, R. Vieira, and S. Teufel. 1997. Resolving 
bridging references in unrestricted text. In R. Mitkov, 
editor, Proc. of the ACL Workshop on Operational 
Factors in Robust Anaphora Resolution, pp. 1-6, 
Madrid. 
M. Poesio, S. Schulte im Walde, and C. Brew. 1998. 
Lexical clustering and definite description interpre- 
tation. In Proc. of the AAAI Spring Synlposiunt on 
Learning for Discourse, Stanford, CA, March. 
M. Pollack, J. Hirschberg, and B. Webber. 1982. User 
participation in the reasoning process of expert sys- 
tem. In Proc. of AAAI-82, pp. 358-361. 
C. L. Sidner. 1979. Towards a computational theo~' 
of definite anaphora comprehension in English dis- 
course. Ph.D. thesis, MIT. 
M. A. Walker. 1996. Limited attention and discourse 
structure. Computational Linguistics, 22(2):255-264. 
B. L. Webber. 1991. Structure and ostension in the inter- 
pretation of discourse deixis. Languageand Cognitive 
Processes, 6(2): 107-135. 
