An Architecture for Anaphora Resolution 
Elaine Rich 
Susann LuperFoy 
MCC, Advanced Computer Architecture Program 
3500 West Balcones Center Drive 
Austin, TX 78759 
ABSTRACT 
In this paper, we describe the pronominal 
anaphora resolution module of Lucy, a portable 
English understanding system. The design of this 
mo;clule was motivated by the observation that, al- 
though there exist many theories of anaphora resolu- 
tion, no one of these theories is complete. Thus we 
have implemented a blackboard-like architecture in 
which individual partial theories can be encoded as 
separate modules that can interact to propose can- 
didate antecedents and to evaluate each other's proposals. 
INTRODUCTION 
The Lucy system (Rich, 1987) is a prototype 
of a portable English front end for knowledge- 
based systems. The major components of Lucy 
are a syntax-based parser (Wittenburg, 1986), a 
semant=c translation system, a pronominal 
anaphora resolution system (which will be 
described in this paper) and a pragmatic proces- 
sor. The parser p..roduces as !ts output a feature 
graph that descnoes the syntacticpropenies ov 
the constituents of the sen(ence. The semantic 
translation system produces as its output a list of 
discourse referents and a set of assertions 
about them. The job of the anaphora resolution 
system is to augment this assertion set with ad- 
ditional assertions that describe coreference relations between discourse referents. Figure 1 
shows the results of semantic processing and 
anaphora resolution for the simple discourse, 
"Dave created a file. He printed it." 
A DISTRIBUTED ARCHITECTURE 
Designing an anaphora resolution system is 
difficult because there exists no single, coherent 
theory upon which to build, even =f we restrict 
our attention to pronominal anaphora (which we 
will do throughout this paper). There do, 
however, exist many partial theories, 1 each of 
which accounts for a subset of the phenomena 
that influence the use and interpretation of 
pronominal anaphora. These partial theories 
range from purely syntactic ones (for example 
the simple rules of number and gender 
agreement) to highly semantic and pragmatic 
ones that account for focusing phenomena. If 
there were a single, complete theory, then it 
might be appropriate to implement it. If there 
were no theories at all, then an ad hoc im- p 
lementation might be the only alternative. But 
ecause there are partial theones but not a com- 
plete one, we have designed an architecture 
(patterned after the idea of a blackboard system 
(Erman, 1981)) that allows for a loosely coupled 
set of modules, each of which handles a subset 
of discourse phenomena by implementing a 
specific partial theory. These modules com- 
municate by proposing candidate antecedents 
and by evaluating each other's proposals. An 
oversight module, called the handler, mediates 
these communications and resolves conflicts 
among the modules. 
All the modules in this system share a com- 
mon representation of the current discourse. 
This representation is called the core discourse 
structure. It includes a list of the discourse 
referents that have so far been introduced. As- 
sociated with each such referent is a set of 
assertions, including syntactic facts about the 
use of the referent (e.g., its number and gender 
and whether or not it is reflexive), semantic facts 
(such as the ones shown in Figure 1), and 
anaphoric usage facts such as coextension rela- 
tions. 
A schematic view of the architecture is 
shown in Figure 2. Each of the ovals in the 
figure represents an implementation of one of 
the partial theories of anaphora. Each of these 
implementations is called a constraint source 
(CS), because each of them is viewed as impos- 
ing a set of consiraints on the choice of an an- 
tecedent for a pronominal referent. 
el: (create el) e2: (pdnt e2) (agent el xt ) (agent e2 x3) 
(object el x2) (object e2 x4) 
xl : (= xl Dave) x3: 
x2: (file x2) x4: 
(a) The Result of Semantic Processing 
1All existing theories are partial in the added sense of 
being fallible. That is, even when restricted to a narrow 
subdomaln of the facts of anaphoric behavior, no account 
fully explains coreference possibilities that arise in context. 
18 
x3: (coextensive x3 xl ) x4: (coextensive x4 x2) 
(b) Assertions Added by Anaphora Resolution 
Figure 1: Processing "Dave created a file. He printed it." 
Cataphora 
Global Focus 
Number Agreement Semantic -" Consistency 
Animacy Disioint Reference 
Figure 2: The Architecture of Anaphora Resolution 
THE STRUCTURE OF A CONSTRAINT 
SOURCE 
Each constraint source in this system is 
composed of a set of four f.unctions (although 
any ot these functions may be a no-op in any 
particular CS). These component functions are 
called at different times during the process of 
anaphor resolution, but they form a logical unit 
since they share knowledge structures that cor- 
respond to the particular partial theory that is implemented by the CS to which they belong. 
The four functions are the following: 
• Modeller - This function maintains the CS's 
local model of the ongoing discourse. Not all 
CS's need such a model. Simple syntactic 
ones like number agreement typically do not 
build a local model since all the information 
they need is available in the core discourse 
structure and in the anaphoric referent being 
resolved. But CS's that describe more global 
phenomena, such as rhetorical structure may 
need access to informat on that is neither lo- 
cal to a particular referent nor contained in the 
core discourse structure. If a local model is 
built, it is built to rest top core on of the shared 
discourse structure. Thecontent of the local 
model is accessible to the other functions of 
the same CS but not to anyone else. 
• Constraint poster - This function posts con- 
straints that describe interactions among 
anaphora within a sentence. .These con- 
stra~nts are then treated in exactly the same 
way as are semantic assertions about the 
referents. The algorithm for resolving 
anaphora (which will be described below) ap- 
I~lies to a single anaphoric referent at a time. 
But, in some sentences, there are interactions 
among referents that make it impossible to 
treat them completely separately. By posting 
interaction constraints before processing in- dividual anaphoric referents, we maintain the 
ability to treat the referents separately but 
also to guarantee that the complete inter- 
pretation for the sentence in which they occur 
will be consistent. As an example of this ~ 
henomenon, consider the sentence, "He saw 
im." Disjoint reference could post a con- 
straint that he and him cannot co-refer before 
any attempt is made to find the antecedent for 
either of them. Most simple constraint 
sources do nothing when their constraint 
posters are called. 
• Proposer - This function takes as input an 
anaphoric referent and returns as its output a 
list of candidate antecedents, each with an associated score (which will be described 
later). For example, recency proposes all 
referents in the two most recent previous sen- 
tences, as well as any in the current sentence 
that occur to the left of the referent that is 
being resolved. Some CS's, such as number 
agreement, never propose candidates. 
• Evaluator - This function takes as input an 
anaphor and a candidate antecedent for that 
anaphor. The evaluator returns a score in- 
dicating strength of support for that candidate 
as antecedent to the anaphor. The returned 
score is based solely on the information avail- 
able to the CS that ~s doing the evaluation. It 
is left to the handler, which invokes the 
various CS's, to combine scores from different evaluators and to resolve conflicts. Although 
every CS must be able to respond whenever 
its evaluator is called, there is a score that will 
be interpreted to mean, "1 have no opinion." 
(See below.) 
THE ANAPHORA RESOLUTION 
PROCEDURE 
In the current implementation of Lucy, 
anaphora resolution occurs once semantic inter- 
19 
pretation is complete. 2 The procedure 
resolve-anaphors does the following for each 
sentence that is to be understood: 
1. Update the core discourse structure with 
the results of syntactic and semantic 
processing of the current sentence. 
2. For each CS in the system, invoke the 
modeller and the constraint poster. 
3. For each anaphor in the sentence do: 
a. Invoke the anaphora resolution hand- 
ler, which will in turn invoke the 
PrrOposers and evaluators for all CS's. he output of the handler is a list of 
ordered pairings of possible antece- 
dents with their overall scores. 
b. Invoke select-best-antecedent, which 
will decide whether there is enough 
evidence to make a coextension 
assertion and, if so, will add the asser- 
tion to the core discourse structure 
and to both anaphor and antecedent 
referents. 
The anaphora resolution handler takes two 
inputs: an ordered list of available CS's and the 
anaphoric referent that needs to be resolved. 
The handler exploits a local data structure called 
CANDIDATES, which is reinitialized for each 
anaphoric referent that is processed and which 
contains a list of all the antecedents that have 
been proposed for that referent. This makes it 
possible to avoid considering any given can- 
didate more than once, no matter how often it is 
proposed. The handler proceeds as follows: 
1. For each CS(i) do: 
a. Invoke CS(i)'s proposer, which will 
return a list (possibly empty) of can- 
didate antecedents, each with an as- 
sociated initial score. 
b. For each candidate C that is not al- 
ready on CANDIDATES do: 
i. Add C to CANDIDATES. 
ii. While running score of C is 
above threshold, for each CS(j) 
where j ~ i do: 
1. Pass C to the evaluator to 
get a score. 2. Update running score for C. 
Although these algorithms do not care, in 
some sense, what list of constraint sources they 
are given, their success depends on having a 
set of constraint sources thatcover the range of 
phenomena that occur in the discourse that 
must be processed. Also, the efficiency with 
which they reach a conclusion depends on the 
order in which the CS's are invoked. In Lucy, 
the recency constraint source is invoked first. 
The correct antecedent is almost always among 
the candidates recency proposes since it proposes liberally. To aidefficiency, the recency 
CS is followed immediately by the simple syn- 
2In future releases, a more flexible control structure will 
be exploited. 
tactic filtering CS's (number, gender and 
animacy agreement) wh=ch, with little effort and 
high certainty, are able to eliminate most can- 
didates from the list that must be evaluated by 
more complex CS's. 
The select-best-antecedent procedure ap- 
plies to the final list of candidate antecedents. It 
must decide whether there is sufficient infor- 
mation on which to base a coreference asser- 
tion. If there is exactly one candidate with the 
highest rating and if the difference between that 
ratingand the second highest rating exceeds the 
threshold 5 (a system parameter currently set at 
0.5), then a coreference assertion can be 
made. 3 If one candidate is not cleady best, 
however, two actions are possible. One ~s to do 
nothing (i.e., post no coreference constraint) and 
thus essentially to produce a partial interpreta- 
tion of the input sentence. This may be accept- 
able, depending on the use to which the 
sentence's interpretation is to be put. For ex- 
ample, it may be possible to wait until sub- 
sequent sentences are processed to see if they 
yield the disambiguating information. If waiting 
~s unacceptable, however, the other available 
action is to query the user who input the sen- 
tence. This option is available since Lucy is 
being designed to serve as an interactive 
Enghsh front end system; if this same approach 
were to be used during text comprehension, 
some alternative, such as choose a candidate 
and be prepared to back up if necessary, would 
be requ=red instead. 
THE CANDIDATE SCORING 
PROCEDURE 
As we just saw, the final selection of an an- 
tecedent from among the set of candidate 
referents depends on the combined score that is 
attached to each candidate as a result of the 
examination of the candidate by the entire set of 
constraint sources. Thus the design of the scor- 
ing procedure has an important effect on the 
outcome of the resolution process. Our first im- 
plementation of this procedure exploited a single 
score, a number in the range -5 to +5. Each CS gave each candidate a score and the handler 
averaged the individual scores to form a com- 
posite score. The major drawback of this simple 
scheme is that there is no way for a CS to say, "1 
have no opinion," since any score it gives neces- 
sarily changes the composite score. There is 
also no way to say, "1 have an opinion and here it is, (and I'm very)/(but rm not at all) confident 
of it." As a result, the system was highly un- stable. It could be tuned to perform fa=rly well, 
3In the current implementation, the value of 8 does not 
matter very much, since the available CS's provide only very 
weak preferences or absolute filtering. Future CS's will 
exploit more domain knowledge to provide more accurate 
preferences. 
2,0 
but whenever a new CS was added or if a CS 
was changed even slightly, the whole system 
had to be retuned. 
To remedy these problems, we now use a 
scoring procedure in which each CS provides 
both a score and a confidence measure. The 
score is a number in the range -5 to +5, the 
confidence a number in the range 0 to 1. The 
function that combines a set of n 
(score, confidence) pairs is 
It 
~, score(Oxconfidence(O 
i=1 running score = 
II 
confidence(t) 
i=1 
This function computes an average that is 
weighted not by the number of distinct scores 
but by the total confidence expressed for the 
scores. Any CS that wishes to assert no opinion 
can now do so by giving a confidence of 0 to its 
opinion, which will then have no effect on a 
candidate's running score. 
Although, in principle, a constraint source 
may function both as a proposer and as an 
evaluator and it may assign any 
(score, confidence) value it likes to a candidate, 
it turns out that the CS's that have been imple- 
mented so far are much more limited in their 
behavior. Most CS's either propose or evaluate, 
but not both. And there are patterns of 
(score, confidence) values that appear to be par- 
ticularly useful. Each CS that has been imple- 
mented so far falls into one of the following four 
classes: 
• Finite set generators (such as disjoint refer- 
ence when applied to a reflexive pronoun) are 
constraint sources that propose a fixed set of 
candidates. They assign all such candidates 
the same score and that score is a function of 
the number of competing candidates: 
# of candidates (score, confidence) contribution 
to propose of this CS 
1 (5.1) =5 
2 (4, 1) =4 
3 (3, 1) =3 
These CS's never evaluate (i.e., when asked 
to do so, they return a confidence of 0.) 
• Fading infinite set generators (such as 
recency) are constraint sources that could 
keep proposing, perhaps indefinitely, but with 
lower and lower scores. Recency, for ex- 
ample, uses the following scoring: 
Sentence (score, confidence) contribution 
of this CS 
n (current) (1, 0.5) = 2 
n-1 (2, 0.5) = 1 
n-2 (0, 0.5) = 0 
These CS's never evaluate. 
• Filters (such as number and gender 
agreement) are constraint sources that never 
propose candidates. They serve only to filter 
out candidates that fail to meet specific re- 
fquirements (usually syntactic). Filters use the otlowing two assignments when they 
evaluate candidates: 
(score, confidence) contribution 
of this CS 
pass (0 4, 0) = 0 
fail (-5, 0.9) = -5 
These scores have the following effects: 
• pass - Since the confidence level is 0, the 
score does not matter and no change will 
be made to the composite score as a result 
of the evaluator being called. Thus a 
candidate's score is insensitive to the num- 
ber of filter CS's that it passes. 
• fail - The low score with high confidence 
forces the composite score to drop below 
the minimum threshold eliminating this can- 
didate from future consideration. 
• Preferences (such as semantic content 
consistency) are constraint sources that im- 
pose preferences, rather than absolute 
opinions, on a set of candidates ratingeach 
member relative to others in the set. These 
constraint sources may use the full range of 
(score, confidence) values. 
Although the scoring scheme we have just 
described exploits more knowledge about a 
CS's opinion than did our first, simpler one, it is 
not perfect. It can suffer from the usual 
problems that arise when numbers are used to 
represent uncertainty. Future implementations 
of this system may move more in the direction ot 
symbolic justifications (as used, for example, in 
(Cohen, 1985)) if they appear to be necessary. 
CONSTRAINT SOURCE EXAMPLES 
In this section, we describe the constraint 
sources that have been implemented in Lucy as 
well as some (preceded by an asterisk) that are 
envisioned but not yet implemented. 
Recency, whose function is to propose 
referents that have occurred in the recently 
preceding discourse. Recency has no opinion to 
offer on anyone else's proposals. 
Number Agreement, which knows that sin- 
gular pronouns must refer to singular things and 
plural pronouns must refer to plural things. 
Number Agreement does not propose antece- 
dents; instead it serves only as a filter on can- 
didates that are proposed by other CS's. 
Gender Agreement, which knows that any 
pronoun that is marked for gender can refer only 
to something of the same gender as itself. 
Gender serves only as a filter =n the current im- 
plementation. 
Animaoy, which knows that neuter pronouns 
refer to inanimate things, while masculine and 
4When the confidence rating is 0, the score is arbitrary 
given the equation for running score values. 
21 
feminine pronouns must refer to animate things, 
usually people. Animacy functions only as a fil- 
ter. 
Disjoint Reference, which knows about 
structure-based coreference restrictions that 
apply to reflexive and to nonreflexive pronouns 
(as described in theories such as (Reinhart, 
1983)). Disjoint Reference proposes antece- 
dents for reflexive pronouns (as, for example, in 
a sentence like, "John saw himself.") For non- 
reflexive pronouns, it serves as a filter, eliminat- 
ing, for example, John as the antecedent of him 
in the sentence, "John saw him." 
Semantic Type Consistency, which func- 
tions as a filter and constrains antecedents to 
only those referents that satisfy the type con- 
straints imposed by the semantic interpretation 
of the rest of the sentence. For example, con- 
sider the discourse, "The system created an er- 
ror log. It printed it." Assume that the semantic 
interpretation of print imposes the following type 
constraints on its arguments: 
agent: human v computer 
object : information-structure 
Then this CS will reject an error log as the an- 
tecedent of the first occurrence of it, assuming 
that the type hierarchy does not include log as a 
subclass of either human or computer. Further, 
this CS will reject the system as the antecedent 
of the second occurrence of it, assuming that the 
type hierarchy does not include system as a 
subclass of information-structure. 
Global Focus, which knows about objects 
that are globally salient throughout a discourse. 
In the current implementation, global Focus acts 
only when the anaphor being considered is it. In 
that case, it proposes as antecedents all 
referents that are in global focus. (Empirical evidence in support of this strategy is presented 
in (Guindon, 1986).) In the current implemen- 
tation, the target system to which the English 
front end is attached is assumed always to be in 
global focus. 
Cataphora, which knows about a class of 
syntactic constructions in which a pronoun can 
preceed the full lexical NP with which it corefers. 
This CS will propose John as a candidate an- 
tecedent for he in the sentence When he is 
happy, John sings. Cataphora acts as a gener- 
ator and will never reject the proposal of another 
CS. 
*Logical accessibility, which knows about 
the constraints that are imposed on the acces- sibility of referents as a function of their embed- 
ding within logical structures such as quantifiers 
and negation (Kamp, 1981). Logical accessibilty 
functions only as a filter. It rules out, for ex- 
ample, a donkey as the antecedent for it in the 
sentence, "If a farmer doesn't own a donkey, he 
beats it," unless a donkey is interpreted as 
having wide scope over the sentence (i.e., "If 
there is a donkey such that the farmer doesn't 
own it then he beats it.") 
*Semantic content consistency, which ex- 
ploits semantic knowledge about context de- 
pendent phenomena as opposed to simply ap- 
plying static type constraints. The boundary be- 
tween this CS and semantic type consistency is 
clearly fuzzy in general and depends in any par- 
ticular case on the structure of the type hierar- chy that is being used. The key difference be- 
tween the CS's, though, is that accessing a type 
hierarchy is fast, whereas there are cases in 
which th~s CS will have to do arbitrary reasoning. 
*Local Focus, which tracks objects that are 
locally in focus in the discourse. This is the 
phenomenon that is addressed by theories such 
as focus spaces (Grosz, 1977) and centering 
(Grosz, 1986, Brennan, 1987). 
*Rhetorical Structure, which segments and 
organizes the discourse as a set of plans for fulfilling conversational goals. This is the 
phenomenon that is addressed by theories such 
as (Hobbs, 1985). 
*Set generation, which creates set-level 
referents that can serve as antecedents for 
plural pronouns. For example, this CS could 
propose Mary and Sue as the antedecent for they 
in the discourse, "Mary picked up Sue. 
They went to the movies." 
*Generic They, which knows about salient 
individuals and groups, and proposes them as 
antecedents for occurrences of the pronoun they 
in sentences such as, "Why don't they ever fix 
the roads?" 
This list is intended to provide an example of 
the range of phenomena that can be combined 
us!ng .the.architecture we have .described.. It is 
not m~enaed to oe a aefinidve ~is\[ oT constraint 
sources. In fact, the architecture allows for more 
than one implementation (i.e., CS) of a given 
theory or more than one theory (and associated implementations) of a given phenomenon. This 
redundancy can be useful, for example, as way 
of comparing the effectiveness of competing constraint sources within a complete anaphora 
resolution system. 
DEBUGGING WITHIN 
ARCHITECTURE 
NOW that the above architecture has been 
implemented, further development of the system 
consists primarily of additions to the set oT con- 
straint sources and adjustments to score and 
confidence assignments. During a test run the 
developer needs to know which referents are 
being recognized as anaphors, which CS's get 
consulted and in what order and, most impor- 
tantly, what effect each CS has on the overall 
rating received by each proposed antecedent. Our tracing tools will display this information for 
each anaphor processed by the handler in the 
following form. First the name of each proposer 
as it is called and the list of candidates and 
ratings it returns are displayed in the tracing win- 
dow. Then for each of these candidates in turn, 
the name of every evaluator appears as it is 
2,2, 
SAMPLE DISCOURSE: 
Sen:ence z : Jon created a file for himself. 
U-I xl el x2 x3 
TRA CE OUTPUT: 
InvokJ'ng anaphor handler wi~ anaphor: #<X-6> 
Invoking proposer for RECENCY 
Possible antecedents proposed: 
(X-3:1) (X-2:1) (E-1:1) cx-, (Xl (u-1 :1) 
Composite ratings for candidate X-3 
After polling evaluator of TYPE rating is 1.0 
After polling evaluator of IDENTITY ralmg is 1.0 
After polling evaluator of GLOBAL-FOCUS rating is 1.0 
After polling evaluator of ANIMACY rating is -2.857143 
Composite 
After potling 
After polling 
After polling 
After polling 
After polling 
After polling 
After polling 
.After polling 
ratinga for candidate X-2 
evaluator of TYPE rating is 1.0 
evaluator of IDENTITY rating is 1,0 
evaluator of GLOBAL-FOCUS rating is 1.0 
evaluator of ANIMACY rating is 1.0 
evaluator of DISJOINT-REFERENCE radng is 1.0 
evaiuator of GENDER rating is 1.0 
evaluator of NUMBER rating is 1.0 
evaluator of CATAPHORA ra~ng is 1.0 
Composite ratings for candidate E-1 
After polling evaluator of TYPE rating is -2.857143 
Invoking proposer for GENDER 
Possible antecedents proposed: none 
Invoking proposer for NUMBER 
Possible antecedents proposed: none 
Invoking proposer for CATAPHORA 
Possible antecedents proposed: none 
Final candidate ratings: 
X-3 -2.857143 
X-2 1.0 
E-1 -2.857143 
X-1 -2.8571 43 
U-1 -2.857143 
E-5 -2.5 
X-4 -2.5 
X-9 -2.5 
Sentence 2: He sent it to Carl and Dave. 
U-2 x4 e5 x6 x7 x8 
I I 
x9 
ANNOTATIONS: 
begin resolving anaphor "it" 
list of candidates and their initial scores 
evaluate first candidate, "himself" 
animate candidate cannot corefer with "it" 
next candidate to be evaluated is "file" 
no CS evaluator wants to reject or suppor~ 
this candidate 
"file" survives with orignal score of !.0 
next candidate is the event "create" 
event referent filtered out due to type mismatch 
(Recency's remaining candidates get filtered) 
begin calling other proposers and evaluate any 
new referents they introduce as candidates 
filtering CS's have nothing Co pIopose 
Cataphora and Disjoint Reference propose nothing 
new for this anaphor in this discourse 
unordered list of all candidates ever proposed 
all candidates except "file" have been rejected 
minimum score threshold currently set at -2.50 
difference between "file" x-2 and first runner-up 
(4.5) exceeds delta currently set at 0.5 
Figure 3: Tracing the Anaphora Resolution Process 
Z3 
called followed by the effect of that evaluator's 
response on the running score for the candidate referent. At the end of processing for each 
anaphor the list of all candidates ever proposed 
and their composite ratings is displayed. Figure 3 shows an example of the use of the tracing 
tools. 
CONCLUSION 
In this paper, we have described an architec- ture for pronominal anaphora resolution that al- 
lows implementations of partial theories of anaphora to be combined into a complete sys- 
tem, and we have illustrated an implementabon of such a system. This architecture makes no 
commitment on the question of what theories should be used or how conflicts among the 
theories should be resolved. In this respect, it 
differs from other proposals (such as (Hobbs, 1978)) in which a specific strategy for applying 
knowledge is encoded into the control structure of the system. As a result of its loose structure, 
this architecture supports the empirical inves- 
tigation of the effectiveness of competing theories and their implementations within a com- 
plete anaphora resolution system. 
One interesting comment that can be made about this architecture is its similarity to architec- 
tures that have been used to perform other parts 
of the natural language understanding task. For 
example, TEAM (-Grosz, 1987) uses a similar ar- 
chitecture and a set of critics to .perform.quan- drier scope assignment. The criucs Tuncfion in 
much the same way CS's do. And, like CS's, 
there are classes of critics. For example, some are pure filters. Others impose preferences on 
the set of candidate interpretations. 
ACKNOWLEDG EMENTS 
We would like to thank Nicholas Asher, Kent 
Wittenburg, Dave Wroblewski, Jim Barnett, Jon 
Schlossberg, and Carl Weir for many discus- sions about this architecture. We wouldalso like 
to thank Carl Weir for his contribution to the im- plementation of this system. 

REFERENCES 

Brennan, S. E., M. W. Friedman, & C. J. Pollard. 
(1987). A Centering Approach to Pronouns. 
Proceedings ACL.

Cohen, R. R. (1985). Heuristic Reasoning about 
Uncertainty: An Artificia/ Intelligence 
Approach. Boston: Pitman Advanced 
Publishing Program. 

Erman, L. D., P. E. London, & S. F. Fickas. 
(1981). The Design and an Example Use of 
Hearsay II1. Proc. IJCAI 7.. 

Grosz, B. J. (1977). The Representation and 
Use of Focus in a System for Understanding 
Dialogs. IJCAI 5.. 

Grosz, B. J., A. K. Joshi, & S. Weinstein. 
(1986). Towards a Computational Theory of 
Discourse Interpretation. 

Grosz, B. J., D. E. Appelt, P. A. Martin, & 
F. C. N. Pereira. (May 1987). TEAM: An 
Experiment in the Design of Transportable 
Natural-Language Interfaces. Artificial 
Intelligence, 32(2), 173-243. 

Guindon, R., P. Sladky, H. Brunner & J. Conner. 
(1986). The Structure of User-Advisor 
Dialogues: Is there Method in their Mad- 
ness? Proceedings of the 24th Meeting of 
the Association for Computational 
Linguistics.

Hobbs,J. R. (1978). Resolving Pronoun 
References. Lingua, 44, 311-338. 

Hobbs, J. R. (1985). On the Coherence and 
Structure of Discourse (Tech. Rep.). 
CSLI-85-37, 

Kamp, H. (1981). A Theory of Truth and 
Semantic Representation. In J. Froenendijk, 
T. Janssen, & M. Stokhof (Eds.), Formal 
Methods in the Study of Language, Part I. 
Amsterdam, The Netherlands: Mathematisch 
Centrum. 

Reinhart, T. (1983). Anaphora and Semantic 
Interpretation. Chicago, Ill.: University of 
Chicago Press. 

Rich, E. A., J. Barnett, K. Wittenburg & 
D. Wroblewski. (1987). Ambiguity Procras- 
tination. Proceedings AAA187. . 

Wittenburg, K. (1986). A Parser for Portable NL 
Interfaces Using Graph-Unification-Based 
Grammars. Proceedings AAA186.
