Focusing and Reference Resolution in PUNDIT 
Focusing and Reference Resolution in PUNDIT 
Deborah A. Dahl 
Research and Development. Division 
SDC -- A Burroughs Company 
PO Box 517 
Paoli, PA 19301 
Science Track 
Natural Language 
ABSTRACT 
This paper describes the use of focusing in the PUNDIT text processing system. 1 
Focusing, as discussed by \[Sidner1979\] (as well as the closely related concept of center- 
ing, as discussed by \[Grosz1983\] ), provides a powerful tool for pronoun resolution. 
However, its range of application is actually much more general, in that it can be used 
for several problems in reference resolution. Specifically, in the PUNDIT system, focus- 
ing is used for one-anaphora, elided noun phrases, and certain types of definite and 
indefinite noun phrases, in addition to its use for pronouns. Another important feature 
in the FUNDIT reference resolution system is that tile focusing algorilJnn is based on 
syntactic constituents, rather than on thematic roles, as in Sidner's system. This 
feature is based on considerations arising from the extension of focusing to cover one- 
anaphora. These considerations make syntactic focusing a more accurate predictor of 
the interpretation of one-anaphoric noun phrases without decreasing the accuracy for 
definite pronouns. 
1 This work is supported in part by DARPA under contract N00014-85-C-0012, administered by the Office of Naval 
: Research. 
114 
Focusing and Reference Resolution in PUNDIT 
1. Background 
1.1. Focusing 
Linguistically reduced forms, such as pronouns, are typically used in texts to refer 
to the entity or entities with which the text is most centrally concerned. 2 Thus, keep- 
ing track of this entity, (the topic, of \[Gundel1974\], the focus of \[Sidner1979\], and the 
backward-looking center of \[Grosz1983, Kameyama1985\] ) is clearly of value in the 
interpretation of pronouns. However, while 'pronoun resolution' is generally presented 
as a problem in computational linguistics to which focusing can provide an answer (See 
for example, the discussion in \[Hirst1981\]), it is useful to consider focusing as a prob- 
lem in its own right. By looking at focusing from this perspective, it can be seen that 
its applications are more general than simply finding referents for pronouns. Focusing 
can in fact play a role in the interpretation of several different types of noun phrases. 
In support of this position, I will show how focus is used in the PUNDIT (Prolog 
UNDerstander of Integrated Text) text processing system to interpret a variety of 
forms of anaphoric reference; in particular, pronouns, elided noun phrases, one- 
anaphora, and context-dependent full noun phrase references. 
A second position advocated in this paper is that surface syntactic form can pro- 
vide an accurate guide to determining what entities are in focus. Unlike previous focus- 
ing algorithms, such as that of \[Sidner1979\], which used thematic roles (for example, 
theme, agent, instrument as described in \[Gruber1976\] ), the algorithm used in this 
system relies on surface syntactic structure to deternfine which entities are expected to 
be in focus. The extension of the focusing mechanism to handle one-anaphora has pro- 
vided the major motivation for the choice of syntactic focusing. 
The focusing mechanism in this system consists of two parts--a FocusList, which 
is a list of entities in the order in which they are to be considered as loci, and a focus- 
ing algorithm, which orders the FocusList. The implementation is discussed in detail 
in Section 5. 
1.2- Overview of the PUNDIT System 
I will begin with a brief overview of the PUNDIT system, currently under 
development at SDC. PUNDIT is written in Quintus Prolog 1.5. It is designed to 
integrate syntax, semantics, and discourse k1~owledge in text processing for limited 
domains. The system is implemented as a set of distinct interacting components which 
communicate with each other in clearly specified and restricted ways. 
The syntactic component, Restriction Grammar,\[Hirschman1982, Hirschman1985\], 
performs a top-down parse by interpreting a set of context-free BNF definitions and 
enforcing context-sensitive restrictions associated with the BNF definitions. The gram- 
mar is generally modelled after that developed by the NYU Linguistic String Project 
\[Sager1981\]. Restrictio~s which enforce context-sensitive constraints on the parse are 
associated with the bnf rules 
I am grateful for the helpfnl comments of Lynette Hirschman, Marcia Linebarger, Martha Palmer, and Rebecca Schiffman 
on this paper. John Dowding and \[3onnie Webber also provided useful comments and suggestions on an earlier version. 
115 
Focusing and Reference Resolution in PUNDIT 
Some semantic filtering of the parse is done at the noun phrase level. That is, 
alter a noun phrase is parsed, it is passed t~o the noun phrase semantics component, 
which determines if there is an acceptable semantics associated with that parse. If the 
noun phrase is acceptable, the semantics component constructs a semantic represents- 
Lion. If the noun phrase is not semantically acceptable, another parse is sought. 
At the conclusion of parsing, the sentence-level semantic interpreter is called. This 
interpreter is based on Pahner's Inference Driven Semantic Analysis system, \[Pal- 
mer1985\], which analyzes verbs into their component meanings and fills their thematic 
roles. In the process of filling a thematic role the semantic analyzer calls reference reso- 
lution for a specific syntactic constituent in order to find a referent to fill the role. 
Reference resolution instantiates the referent, and adds to the discourse representation 
any information inferred during reference resolution. 
Domain-specific information is available for both the noun phrase and clause level 
semantic components through the knowledge base. The domain currently being 
modelled by SDC is that of computer maintenance reports. Currently the knowledge 
base is implemented as a semantic net containing a part-whole hierarchy and an iua 
hierarchy of the components and entities in the application domain. 
Following the semantic analysis, a discourse component is called which updates 
the discourse representation to include the information from the current sentence and 
which runs the focusing algorithm. 
2. Uses of Focusing 
Focusing is used ill four places in PUNDIT -- for definite pronouns, for elided 
noun phrases, for one-anaphora, and for implicit associates. 
As stated above, reference resolution is called by the semantic interpreter when it 
is in the process of filling a thematic role. Reference resolution proposes a referent for 
the constituent associated with that role. For example, if the verb is replace and the 
semantic interpreter is filling the role of agent, reference resolution would be called 
for the surface syntactic subject. After a proposed referent is chosen tbr the subject, 
any specitic selectional restrictions on the agent of replace (such as the constraint that 
the agent has to be a human being) are checked. If the proposed referent fails selec- 
tion, backtracking into reference resolution occurs and another referent is selected. 
Cooperation between reference resolution and the semantic interpreter is discussed in 
detail in \[Pahner1986\]. The semantic interpreter itself is discussed in \[Palmer1985\]. 
2.1. Pronouns and Elided Noun Phrases 
Pronoun resolution is done by instantiating the referent of the pronoun to the first 
member of the FocusList unless the instantiation would violate syntactic constraints 
on coreferentiality. 3 (As noted above, if the proposed referent fails selection, 
3 At the moment, the syntactic constraints on coreferentiality used by the system are very simple. If the direct object is 
reflexive it must be instantiated to the same referent, as the subject. Otherwise it must be a different referent. Obviously, as the 
system is extended to cover sentences with more complex structures, a more sophisticated treatment of syntactic constraints on 
116 
Focusing and Reference Resolution in PUNDIT 
backtracking occurs, and 'another referent is chosen.) 
The reference resolution situation in the maintenance texts however, is compli- 
cated by the fact that there are very few overt pronouns. Rather, in contexts where a 
noun phrase would be expected, there is often elision, or a zero-np as in Won't power 
up and Has not failed since Hill's arrival. Zeroes are handled exactly as if they were 
pronouns. The hypothesis that elided noun phrases can be treated in the same way as 
pronouns is consistent with previous claims in \[Gundel1980 \] and \[Kameyama1985\] that 
in languages such as Russian and Japanese, which regularly allow zero-np's, the zero 
corresponds to the focus. If these claims are correct, it is not surprising that in a sub- 
language like that found in the maintenance texts, which also allows zero-np's, the zero 
should correspond to the focus. 
Another kind of pronoun (or zero) also occurs in the maintenance texts, which is 
not associated with the local focus, but is concerned with global aspects of the text. 
For example, the field engineer is a default agent in the maintenance domain, as in 
Thinks problem is in head select area. This is handled by defining default elided 
referents for the domain. The referent is instantiated to one of these if no suitable 
candidate can be found in the FocusList. 
2.2. Implicit Aaaoclatea 
Focttni,,g is also used in the processing of certain full noun phrases, both definite 
and inde\[i,,ite, which involve implicit associates. The term implicit associates refers 
to tile relationship between a disk drive and the motor in examples like The field 
engineer installed a disk drive. The motor failed. It is natural for a human reader to 
infer that the motor is part of the disk drive. In order to capture this intuition, it is 
necessary for the system to relate the motor to the disk drive of which it is part. Rela- 
tionships of this kind have been extensively discussed in the literature on definite refer- 
ence. For example, implicit associates correspond to inferrable entities described by 
\[Prince1981\], the associated use definites of \[Hawkins1978\], and the associated type 
of implicit backwards specification discussed by \[Sidner1979\]. Sidner suggests that 
implicit associates should be found among the entities in focus. Thus, when the system 
encounters a definite noun phrase mentioned for the first time, it sequentially examines 
each member of the FoeusList to determine if it is a possible associate of the current 
noun phrase. The specific association relationships (such as part-whole, object- 
property, and so on) are defined in the knowledge base. 
This mechanism is also used in the processing of certain indefinite noun phrases. 
In every domain, it is claimed, there are certain types of entities which can be 
classified as dependent. By this is meant an entity which is not typically mentioned on 
its own, but which is referred to in connection with another entity, on which it is 
dependent. In the maintenance domain, for example, parts such as keyboards, motors, 
and printed circuit boards are dependent, since when they are mentioned, they are nor- 
mally mentioned as being part of something else, such as a console, disk drive, or 
coindexing using some of the insights of \[Reinhart1976\], and \[Chomsky1981\] will be required. 
117 
Focusing and Reference Resolution in PUNDIT 
printer. 4 In an example like The system is down. The field engineer replaced a bad 
printed circuit, board, it seems clear that a relai,ionship between the printed circuit 
board and the system should be represented. Upon encountering a reference to a 
dependeni, eni,ity like the printed circuii, board, the system looks through the 
FocusList~ 13o determine if any previously mentioned entities can be associated with a 
printed circuii, board, and if so, the relationship is made explicit. If no associate has 
been mentioned, the entity will be associated wii,h a default defined in the knowledge 
base. For example, in the maintenance domain, parts are defined as dependent enti- 
ties, and in the absence of an explicitly mentioned associate, they are represented as 
associated with the system. 
2.3- One-Anaphora 
PUNDIT extends focusing to the analysis of one-anaphora following \[Dahl1984\], 
which claims that focus is central to the interpretation of one-anaphora. Specifically, 
the referent of a one-anaphoric noun phrase (e.g., the blue one, some large ones) is 
claimed to be a member or members of a set which is the focus of the current clause. 
For example, in Installed two disk drives. One failed, the set of two disk drives is 
assumed to be the focus of One failed, and the disk drive I hat failed is a member of 
thai, set. This analysis can be contrasted with thai, of \[ll:,,lliday1976\], which treats 
one-anaphora as a surface syntactic phenomenon, completely distinct from reference. 
It is more consistent with the theoretical discussions of \[1976\], and \[Webber1983\]. 5 
These analyses advocate a discourse-pragmatic I,reatment for both one-anaphora and 
definite pronouns. The main computational advantage of treating one-anaphora as a 
discourse problem is that, since definite pronouns are treated this way, little 
modification is needed to the basic anaphora mechanism to allow it to handle one- 
anaphora. In contrast, an implementation following the account of Halliday and 
Hasan would be much more complex and specitic to one-anaphora. 
The process of reference resolution for one-anaphora occurs in two stages. The 
first stage is resolution of the anaphor, one, and this is the stage that involves focus- 
ing. When the system processes the head noun one, it instantiates it with the 
category of the first set in the FocusList; (disk drive in this example). 6 In other 
words, the referent of the noun phrase must be a member of the previously mentioned 
set of disk drives. The second stage of reference resolution for one-anaphora assigns a 
specific disk drive as the referent of the entire noun phrase, using the same procedures 
that would be used for a full noun phrase, a disk drive. 
The extension of the system to one-anaphora provides the clearest motivation for 
the choice of a syntactic focus in PUNDIT. Before I discuss the kinds of examples 
4 There are exceptions to this generalization. For example, in a sentence like field engineer ordered motor, the motor on 
order is not part of anything else {yet). In PUNDIT, these cases are assumed to depend on the verb meaning. In this example, the 
object of ordered is categorized as nort-specifie, and reference resolution is not called. See \[Palmer1986\] for details. 
s Although not Webber's analysis in \[Webber19781, which advocates an approach similar to Halliday and Hasan's. 
e Currently the only sets in the FocusL|st are those which were explictly mentioned in the text. ttowever, as pointed out 
by \[Dahl1982.\], and \[Webber1983, Dahl1984\], other sets besides those explictly mentioned are available for anaphoric reference. 
These have not yet been added to the system. 
118 
Focusing and Reference Resolution in PUNDIT 
which support this approach, I will briefly describe the relevant part of the focusing 
algorithm based on thematic roles which is proposed by\[Sidner1979\]. After each sen- 
tence, the focusing algorithms order the elements in the sentence in the order in which 
they are to be considered as potential loci in the next sentence. Sidner's ordering and 
that of PUNDIT are compared in Figure 1. 
The idea that surface syntax is important in focusing comes from a suggestion by 
\[Erteschik-Shir1979\], that every sentence has a dominant syntactic constituent, which 
provides a default topic for tile following utterance 7. Intuitively, the dominant consti- 
tuent can be thought of as the one to which tile hearer's attention is primarily drawn. 
Operationally the dominance of a constituent is tested by seeing if a referent with that 
constituent as the antecedent can be cooperatively referred to with an unstressed pro- 
noun in the following sentence. 
The feature of onc-anaphora which motivates the syntactic algorithm is that the 
availability of certain noun phrases as antecedents for onc-anaphora is strongly 
affected by surface word 'order variations which change syntactic relations, but which 
do not affect thematic roles. If thematic roles are crucial for focusing, then this pattern 
would not be observed. 
Consider the following examples: 
(i) A: I'd like to plug in this lamp, but the bookcases blocking are 
outlets. 
the electrical 
\]3: Well, can we move one? 
(2) A: I'd like to plug in this lamp, but the electrical outlets are blocked by the book- 
cases. 
Sidner PUNDIT 
Theme 
Other thematic roles 
Agent 
Verb Phrase 
Sentence 
Direct Object 
Subject 
Objects of Freposltlonal Phrases 
Figure 1: Comparison of Potential Focns Ordering in 
Sidner's System and PUNDIT 
7 As discussed in \[Dah119fi4\] there are problems with Erteschik-Shir's definition of dominance and slightly different definition 
is proposed. However the details of this reformulation do not concern us here. 
119 
Focusing andReference Resolution in PUNDIT 
B: Well, can we move one? 
In (1), rnosl, informants report an initial impression that B is talking about moving 
the electrical outlets. This does not happen for (2). This indicates that the expected 
focus following (1) A is the outlets, while it is the bookcases in (1) B. However, in each 
case, the thematic roles are the same, so an algorithm based on thematic roles would 
predict no difference between (1) and (2). 
Similar examples using definite pronouns do not seem to exhibit the same effect. 
In {3) and (4), they seems to be ambiguous, until world knowledge is brought in. Thus, 
in order to handle definite pronouns alone, either algorithm would be adequate. 
(3) A: 
B: 
(4) A: 
B: 
I'd like to plug in this lamp, but bookcases are blocking the electrical outlets. 
Well, can we move them? 
I'd like to plug in this lamp, but the electrical outlets are blocked by the book- 
cases. 
Well, can we move them? 
(5) and (6)illustrate another example with one-anaphora. In (5) but not in (6), 
the initial interpretation seems to be that a bug has lost its leaves. As in (1) and (2), 
however, the thematic roles are the same, so a thematic-role-based algorithm would 
predict no difference between the sentences. 
(5) The plants are swarming with the bugs. One's already lost all its leaves. 
(6) The bugs are swarming over the plants. One's already lost all its leaves. 
In addition to theoretical considerations, there are a number of obvious practical 
advantages to defining tbcus on constituents rather than on thematic roles. For exam- 
ple, constituents can often be found more reliably than thematic roles. In addition, 
thematic roles hage to be defined individually for each verb. 8 Since thematic roles for 
verbs can vary across domains, defining focus on syntax makes it less domain depen- 
dent, and hence more portable. While in principle focus based on thematic roles does 
not have to be domain-dependent, a general algorithm based on thematic roles would 
have to rely on a a general, domain-neutral specification of all possible thematic roles 
and their behavior in focusing. Until such a specification exists, a thematic-role based 
focusing algorithm must be redefined for each new domain as thedomain requires the 
definition of new thematic roles, and because of this, will continue to be less portable 
than an approach based oll syntax. 
8 Of course, some generalizations can be made about how arguments map to thematic roles. For example, tile basic 
definition of the thematic role theme is that, for a verb of motion, the theme is the argument that moves. More generally, the 
theme is tile argument that is most affected by the action of the verb, and its typical syntactic manifestation is as a direct object 
of a transitive verb, or the subject of an intransitive verb. However, even if these generalizations are accurate, they are no more 
than guidelines for finding tile themes of verbs. The verbs still have to be classified individually. 
120 
Focusing and Reference Resolution in PUNDIT 
3. Implementation 
3.1. The FocusList and CurrentContext 
The data structures that retain information from sentence to sentence in the 
PUNDIT system are the FocuaLiat and the CurrentContext. The FoeuaLiat is 
a list of all the discourse entities which are eligible to be considered as foci, listed in 
the order in which they are to be considered. For example, after a Sentence like The 
field engineer replaced the disk drive, the following FoeuaLiat would be created. 
\[\[ev eat1 \],\[drivel \],\[engineer 111 
The members of the FocusList are unique identifiers that have been assigned to the 
three discourse entities -- the disk drive, the field engineer, and the event. The 
CurrentContext contains the information that has been conveyed by the discourse 
so far. After the example above, the CurrentContext would contain three types of 
information: 
(1) Discourse id's, which represent classifications of entities. For example, 
id(field^engineer,\[engineerl\]) means that \[engineerl\] is a a field engineer. 9 
(2) Facts about part-whole relationships (hasparts). In the example in Figure 2, 
notice that the lack of a representation of time results in both drives being part of 
the system, which they are, but not at the same time. W~,rk to remedy this prob- 
lem is in progress. 
(3) Representations of the events in the discourse. For example, if the event is that of 
a disk drive having been replaced, the representation consists of a unique 
identifier (\[eventl\]), the surface verb (replace(time(_))), and the decomposi- 
tion of the verb with its (known) arguments instantiated 1°. The thematic roles 
involved are object1, the replaced disk drive, objeet2, the replacement disk 
drive, time and instrument which are uninstantiated, and agent, the field 
engineer. (See\[Palmer1986\], for details of this representation). Figure 2 illustrates 
how the Curr.entContext looks after the discourse-initial sentence, The field 
engineer replaced the disk drive. 
3.2. The Focusing Algorithm 
The focusing algorithrn used in this system resembles that of \[Sidner1979\], 
although it does not use the actor focus and uses surface syntax rather than thematic 
roles, as discussed above. The focusing algorithm is illustrated in Figure 3. Removing 
candidates from the FoeuaLiat when they are no longer eligible to be the referents of 
pronouns is not currently done in this system. The conditions determining this have 
not been fully investigated, and since the texts involved are short, few problems are 
created in practice. This problem will be addressed by future research. 
m field'englneer is an example of the representation used in PUNDIT for an idiom. 
,s 8176 is an uninstantiated variable representing the time off the replacement. It appears in several places, such M 
inclnded(objectZ(\[driveZD,tlme(_St7O)), and mlsslng(objectl(\[driveli),time(_8176)). 
121 
Focusing and Reference Resolution in PUNDIT 
ia(aeid ^ e.gi.eer, \[e.gi.eer q), 
id(disk ^drive,\[drivel\]), 
ia(ai~k ^ ari~,\[dri~e2 \]), 
haspart(\[syatemll,\[drivell) , 
ha~part(\[~y~temll,\[driveZ\])\] 
event(Ieve.tl\], 
replace( time( _817 6 ) ), 
Iincluded(obj ectZ (\[drive2l),time(8176)), 
missing( object1( \[drivel \]),tlme( 817 6 ) ), 
u~e(i,~t r u me, t(~ .,, 0 ~ ), exchange(objectX(\[dri~el\]),objectZ(\[driveZ\]),time(_alTa))), 
cau~e(age.t(\[e.gi..~erl\]), 
u.e( i.~trume.. ,;( 84o~ ), 
exchange(o bjeetl(\[drive1\]),objeet2(\[drive2\]),time(8176))))\]) 
Figure 2: CurrentContext after The field engineer replaced the disk drive. 
122 
Focusing and Reference Resolution in PUNDIT 
(1) 
(z) 
First Sentence of a Discourse: 
Establish expected foci for the next sentence (order FocusList): the 
order reflects how likely that constituent is to become the focus of 
the following sentence. 
Sentence 
Direct Object 
Subject 
Objects of Prepositional Phrases 
Subsequent Sentences (update FocusList): 
If there is a pronoun in the current sentence, move the focus to the 
referent of the pronoun. If there is no pronoun, retain the focus 
from the previous sentence. Order the other elements in the sen- 
tence as in (1)- 
Figure 3: The Focusing Algorithm 
4. Summary 
Several interesting research issues are raised by this work. For example, what is 
the source of the focusing algorithm? Is it derivable from theoretical considerations 
about how language is processed by human beings, or is it simply an empirical obser- 
vation about conventions used in particular languages to bring discourse entities into 
prominence? Evidence bearing on this issue would be to what extent the focusing 
mechanism c:.trries over to other, non-related languages. Kameyama's work on 
Japanese sug~(,.sts that there are some similarities across languages. To the extent that 
such similarities exist, it would suggest that the algorithm is derivable from other 
theoretical considerations, and is not simply a reflection of linguistic conventions. 
This paper has described the reference resolution component of PUNDIT, a large 
text understanding system in Prolog. A focusing algorithm based on surface syntactic 
constituents is used in the processing of several different types of reduced reference: 
definite pronouns, onc-anaphora, elided noun phrases, and implicit associates. This 
generality points out the usefulness of treating focusing as a problem in itself rather 
than simply as a tool for pronoun resolution. 
123 
Focuslng and Ret'erence Resolution in \]PUNDIT 
REFERENCES 
\[1976\] 
Jorge Hankamer and Ivan Sag, Deep and Surface Anaphora. Linguistic In- 
quiry '/(3), 1976, pp. 391-428. 
\[Chomsky19811 
Noam Chomsky, Lectures on Government and Binding. Foris Publications, 
Dordrecht, 1981. 
{Dah11982.\] 
Deborah A. Dahl, Discourse SI,ructure and one-anaphora in English, presented 
at the 57th Annual Meeting o\[' ~he Linguistic Society of America, San Diego, 
1982.. 
\[Dah11984\] 
Deborah A. Dahl, The Structure and Function of One-Anaphora in English, 
PhD Thesis; (also published by Indiana University Linguistics Club, 1985), 
University of Minnesota, 1984. 
\[Erl, eschik-Shir 1979\] 
Nomi Erteschik-Shir and Shalom Lappin, Dominance and the Functional Ex- 
planation of Island Phenomena. Theoretical Linguistics, 1979, pp. 41-86. 
\[Grosz1983\] 
Barbara Grosz, Aravind K. Joshi, and Scott Weinstein, Providing a Unified 
Account of Definite Noun Phrases in Discourse. Proceedings of the 21st An- 
nual M eetin 9 of the Association for Computational Linguistics, 1983, pp. 
44-50. 
\[Gruberl.976\] 
Jeffery Gruber, Lexical Structure in Syntax and Semantics. North Holland, 
New York, 1976. 
\[Gunde11974\] 
Jeanette K. Gundel, Role of Topic and Comment in Linguistic Theory, Ph.D. 
thesis, University of Texas at Austin, 1974. 
\[Gunde11980\] 
Jeanette K. Gundel, Zero-NP Anaphora in Russian. Chicago Linguistic So- 
ciety Parasession on Pronouns and Anaphora, 1980. 
\[Halliday1976\] 
Michael A. K. Halliday and Ruqaiya Hasan, Cohesion in English. Longman, 
London, 1976. 
\[Hawkins1978\] 
John A. Hawkins, Definiteness and Indefiniteness. Humanities Press, Atlan- 
tic Highlands, New Jersey, 1978. 
124 
Focusing and Reference Resolution in I'UNDIT 
\[Hirschman1982\] 
L. I-Iirschnmn and K. Puder, Restriction Grammar in Prolog. In Proc. of the 
First International Logic Programming Conference, M. Van 
Caneghem (ed.), Association pour la Diffusion et le Developpement de Prolog, 
Marseilles, 1982, pp. 85-90. 
\[Hirschman1985\] 
L. Hirschman and K. Puder, Restriction Grammar: A Prolog 
Implementation. In Logic Programming and its Applications, D.H.D. Warren 
and M. VanCaneghem (ed.), 1985. 
tHirst1981\] 
Graeme Hirst, Anaphora in Natural Language Understanding. Springer- 
Verlag, New York, 1981. 
\[Kamey area 1985 \] 
Megumi Kameyama, Zero Anaphora: The Case of Japanese, Ph.D. thesis, 
Stan|brd University, 1985. 
\[Pahner1985\] 
Martha S. Pahner, Driving Semantics for a Limited Domain, Ph.D. thesis, 
University of Edinburgh, 1985. 
\[Palmer1986\] 
Martha S. Palmer, Deborah A. Dahl, Rebecca J. Schiffman, Lynette Hirsch- 
man, Marcia Linebarger, and John Dowding, Recovering Implicit Information, 
to be presented at the 24th Annual Meeting of the Association for Computa- 
tional Linguistics, Columbia University, New York, August 1986. 
\[Prince1981\] 
Ellen F. Prince, Toward a Taxonomy of Given-New Information. In Radical 
Pragmatics, Peter Cole (ed.), Academic Press, New York, 1981. 
\[Reinhart1976\] 
Tango Reinhart, The Syntactic Domain of Anaphora, Ph.D. thesis, Mas- 
sachusetts Institute of Technology, 1976. 
\[Sager1981\] 
N. Sager, Natural Language Information Processing: A Computer Gram- 
mar of English and Its Applications. Addison-Wesley, Reading, Mass., 1981. 
\[Sidner1979\] 
Candace Lee Sidner, Towards a Computational Theory of Definite Anaphora 
Comprehension in English Discourse, MIT-AI TR-537, Cambridge, MA, 
1979. 
\[Webber1978\] 
Bonnie Lynn Webber, A Formal Approach to Discourse Anaphora. Garland, 
New York, 1978. 
125 
Focusing and Reference Resolution in PUNDIT 
\[Webber19831 
Bonnie Lynn Webber, So What Can We Talk About Now?. In Computational 
Models of Discourse, Michael Brady and Robert C. Berwick (ed.), 1983. 
126 
