A PRAGMATICS-BASED APPROACH TO ELLIPSIS RESOLUTION 
Sandra Carberry 
Department of Computer and Information Sciences 
University of Delaware 
Newark, DE 19716 
Intersentential elliptical utterances occur frequently during information-seeking dialogs in task do- 
mains. This paper presents a pragmatics-based framework for interpreting such utterances. Discourse 
expectations and focusing heuristics are used to facilitate recognition of an information-seeker's intent 
in uttering an elliptical fragment. The ellipsis is comprehended by identifying both the aspect of the 
information-seeker's task-related plan highlighted by the fragment and the conversational discourse 
goal fulfilled by the utterance. The contribution of this approach is its consideration of pragmatic 
information, including discourse content and conversational goals, rather than just the precise 
representation of the preceding utterance. 
1 INTRODUCTION 
Incomplete utterances are common in communication 
between humans. They range from sentences that fail to 
include all requisite semantic information to syntacti- 
cally incomplete sentence fragments. In many cases, 
these utterances cannot be understood in isolation, but 
must be interpreted within the established context. 
Precisely how this should be done is a difficult problem 
for natural language systems. 
One might suggest that the problem be avoided in 
man-machine communication by training human users 
to employ only syntactically and semantically complete 
utterances. However, this does not appear to be feasi- 
ble, as demonstrated in an empirical study conducted by 
Carbonell (1983) in which it was shown that human 
users find it easy to avoid complex syntactic structures 
but difficult to avoid incomplete utterances. 
Even if it were possible to train users to avoid 
incomplete utterances, these restrictions would be un- 
desirable. Constraining man-machine communication to 
only a subset of the utterances normally employed by 
humans would force users to give less attention to their 
problem solving goals in order to concentrate more on 
the preciseness of their utterances. In addition, it ap- 
pears that fragmentary utterances are not merely a 
result of sloppy communication. Although every utter- 
ance has a discourse goal (a conversational or commu- 
nicative goal such as answering a question or seeking 
clarification), elliptical fragments are often used to 
accomplish discourse goals that would require more 
effort to convey with a complete sentence. For exam- 
ple, in the following dialog ~, Speaker 2's fragment 
expresses doubt about the proposition stated by 
Speaker 1. 
Example 1 
Speaker 1: "The Korean jet shot down by the Soviets 
was a spy plane." 
Speaker 2: "With 269 people on board?" 
However, a complete sentence such as 
Speaker 2: "Was the Korean jet shot down by the 
Soviets a spy plane with 269 people on board?" 
fails to adequately communicate the doubt conveyed by 
the previous fragment. Only a more complex sentence 
that marks the discourse goal, such as 
Speaker 2: "How can you think that the Korean jet 
shot down by the Soviets was a spy plane, when it 
had 269 people on board?" 
will accomplish this objective. 
Thus a robust natural language interface must handle 
the kinds of incomplete utterances normally used by 
humans. To do otherwise is to prohibit communication 
that humans regard as natural, and therefore detract 
from their ability to communicate as effectively with 
machines as they do with one another. 
Contextual ellipsis in dialog is the use of a sentence 
fragment (a syntactically incomplete utterance), along 
with the context established by the preceding dialog, to 
communicate a complete thought and accomplish a 
Copyright 1989 by the Association for Computational Linguistics. Permission to copy without fee all or part of this material is granted provided 
that the copies are not made for direct commercial advantage and the CL reference and this copyright notice are included on the first page. To 
copy otherwise, or to republish, requires a fee and/or specific permission. 
0362-613X/89/010075-96-$ 03.00 
Computational Linguistics, Volume 15, Number 2, June 1989 75 
Sandra Carberry A Pragmatics-Based Approach to Ellipsis Resolution 
speech act 2. Such fragments are often referred to as 
intersentential ellipsis since they appear between sen- 
tences in a dialog. It can be seen, however, that 
understanding intersentential ellipsis often depends 
more on pragmatic knowledge, such as the inferred 
task-related plan and discourse goals motivating the 
speaker, than on the syntactic structure or semantic 
content of preceding utterances. This is illustrated by 
Examples 2 and 3, in which Speaker l's communicated 
goal and the relevant plans for accomplishing it play a 
major role in understanding the elliptical fragment. 
Example 2 
Speaker 1: "I want to cash this clheck. Small bills 
only, please." 
Example 33 
Speaker 1: "Who are the candidates for program- 
ming consultants?" 
Speaker 2: "Mary Smith, Bob Jones, and Ann Doe 
have applied for the job." 
Speaker 1: "Tom's recommendation?" 
Speaker 2: "He thinks Bob Jones and Ann Doe have 
the necessary background and should be invited 
for an interview." 
In Example 2, Speaker l's goal is to cash a check, and 
relevant plans can include a subplan for getting a 
particular distribution of paper money; in Example 3, 
Speaker l's goal is hiring programming consultants, 
perhaps for an introductory programming course, and 
the relevant plans for doing this include a subplan for 
identifying the best applicant. 
Previous research on understanding intersentential 
ellipsis has emphasized syntactic and semantic strate- 
gies (Hendrix et al. 1978; Waltz 1978; Weischedel and 
Sondheimer 1982; Carbonell 1983), but the contribu- 
tions of the speaker's plans and goals to the interpreta- 
tion of ellipsis has hitherto been inadequately explored. 
One objective of our research has been to investigate a 
plan-based framework for understanding intersentential 
ellipsis that occurs in task-oriented dialogs. Our work 
shows that elliptical fragments can be viewed as high- 
lighting aspects of the information-seeker's task-related 
plan, with the focus of attention in the plan providing 
the context in which the fragment should be under- 
stood. 
But identifying the aspect of the speaker's plan 
highlighted by the ellipsis is not enough to understand 
the utterance. As shown by Grosz (1979), a speaker may 
pursue several different kinds of goals with a single 
utterance. For example, an information-seeker has the 
long-term goal of performing a task comprised of sub- 
tasks; this hierarchical structure produces a set of 
task-related goals. The information seeker is attempting 
to construct a plan for his underlying task, resulting in 
plan formation goals. In addition, each utterance pur- 
sues a more immediate discourse goal, such as request- 
ing information or seeking clarification. As has been 
shown by several researchers (Mann et al. 1977; Reich- 
76 
man 1978; Pollack et al. 1982), one must determine not 
only the element of the task being addressed by an 
utterance, but also the discourse goal being fulfilled by 
it. Previous research on understanding ellipsis has ig- 
nored this aspect. Thus a second objective of our 
research has been to formulate a strategy for recogniz- 
ing discourse goals communicated by elliptical frag- 
ments. Our work indicates that mutual beliefs and 
expectations developed from the preceding dialog play a 
major role in identifying the intent behind intersenten- 
tial ellipsis. 
Our processing framework coordinates many knowl- 
edge sources, including discourse expectations, inferred 
beliefs, the information-seeker's inferred task-related 
plan, and focusing strategies, to produce a rich inter- 
pretation of ellipsis. It is the first interpretation strategy 
to address the problem of identifying the discourse goal 
accomplished via an intersentential elliptical fragment. 
As a result, this pragmatics-based framework under- 
stands elliptical fragments that other systems cannot 
handle. 
2 STRATEGIES FOR INTERPRETING INTERSENTENTIAL 
ELLIPSIS 
As illustrated by Examples 1, 2, and 3, intersentential 
elliptical fragments cannot be understood in isolation 
from the context in which they occur. Therefore a 
strategy for interpreting such fragments must rely on 
knowledge obtained from sources other than the frag- 
ment itself. Three possibilities exist: the syntactic form 
of preceding utterances, the semantic representation of 
preceding utterances, and expectations gleaned from 
understanding the preceding discourse. 
A strategy using the first kind of knowledge might 
perform two transformations on the syntactic form of a 
preceding utterance. 
1. a substitution/expansion operation in which the 
elliptical fragment is substituted for a syntactic 
constituent in the preceding utterance, or the 
syntactic representation of the preceding utter- 
ance is expanded to accommodate the fragment. 
2. a transformation operation that maps questions 
into answers, statements into questions, etc. 
Syntactic strategies are exemplified by the work in 
Hendrix et al. (1978) and Weischedel et al. (1982). 
Howe, ver, many examples of ellipsis cannot be handled 
by syntactic strategies. Consider again the dialog pre- 
sented in Example 2. No transformation or modification 
of Speaker l's first statement will produce an utterance 
that represents the intended meaning of Speaker l's 
elliptical fragment. 
The second potential strategy uses the semantic 
representation of the last utterance as a pattern that 
suggests slots for which an elliptical fragment may 
provide a filler or a substitution. Semantic strategies are 
presented in Waltz (1978) and Carbonell (1983). How- 
ever, semantic strategies require an extensive case 
Comlmtatioual Linguistics, Volume 15, Number 2, June 1989 
Sandra Carberry A Pragmatics-Based Approach to Ellipsis Resolution 
frame representation and are inadequate for handling 
elliptical fragments that rely on an assumed communi- 
cation of the underlying task. For example, it is improb- 
able that the case frame for programming consultant in 
Example 3 would contain slots for the recommendations 
of individual faculty. 
The third potential strategy utilizes pragmatic knowl- 
edge, such as a model of the information-seeker's 
inferred task-related plan and expected discourse goals. 
The power of this approach is its reliance on knowledge 
gleaned from the dialog, including discourse content 
and conversational goals, rather than on precise repre- 
sentations of the preceding utterances alone. 
Allen (Allen et al. 1980) addressed the problem of 
understanding sentence fragments occurring at the out- 
set of a dialog, and was the first to relate ellipsis 
processing to the domain-dependent plan underlying a 
speaker's utterance. Allen viewed the speaker's utter- 
ance as part of a plan that the speaker had constructed 
and was executing to accomplish his overall task-related 
goals (in this case, to meet or board a train). To interpret 
elliptical fragments, Allen first constructed a set of 
possible surface speech act representations for the 
elliptical fragment, limited by syntactic clues appearing 
in the fragment; he then attempted to infer the speaker's 
goal-related plan that included an execution of the 
observed utterance. A part of this inference process 
involved determining which of the partially constructed 
plans, connecting the task-related goals of meeting or 
boarding a train and the observed utterance, were 
reasonable given the knowledge and mutual beliefs of 
the speaker and hearer. Allen selected the surface 
speech act that produced the most reasonable inferred 
plan as the correct interpretation. 
Allen noted that the speaker's fragment must identify 
the subgoals that the speaker is pursuing, but claimed 
that in very restricted domains, identifying the speak- 
er's overall goal from the utterance was sufficient to 
identify the appropriate response in terms of the obsta- 
cles present in such a plan. For isolated sentence 
fragments in his restricted domain of train arrivals and 
departures, Allen's interpretation strategy worked well. 
To understand elliptical fragments during a dialog in a 
more complex domain, it is necessary to identify the 
particular aspect of the speaker's overall task-related 
plan addressed by the fragment and recognize the 
discourse or communicative goal being pursued. 
More recently, Litman (1986) developed a plan-based 
theory that incorporates both domain plans and meta- 
plans to understand utterances. Similar to Allen, Lit- 
man used syntactic clues in the utterance to postulate 
speech acts "request" or "inform" that an elliptical 
fragment was intended to accomplish, and she then 
attempted to construct an inference path from a postu- 
lated speech act to the plan inferred for the speaker. 
However, it appears that Litman's system would handle 
an elliptical fragment such as 
Computational Linguistics, Volume 15, Number 2, June 1989 
"With 269 people on board?" 
from Example 1 in much the same way as the complete 
sentence 
"Did the Korean jet have 269 people on board?" 
As such, Litman's strategy would be unable to recog- 
nize the surprise and doubt conveyed by the elliptical 
fragment in Example 1. In addition, Litman's metaplans 
(such as Introduce-Plan, Continue-Plan, and Modify- 
Plan) are more like plan construction metaplans than 
metaplans representing communicative goals such as 
expressing surprise or seeking clarification of a ques- 
tion. We contend that recognition of the speaker's 
discourse or communicative goal must be an integral 
part of any strategy that really understands ellipsis. 
In addition to the syntactic, semantic, and plan-based 
strategies, a few other heuristics have been devised. 
Carbonell (1983) used rules that suggested a set of 
expected user utterances and related elliptical frag- 
ments to these expected patterns. For example, if the 
system asked the user whether a particular value should 
be used as the filler of a slot in a case frame, the system 
then expected the user's utterance to contain a confir- 
mation or disconfirmation pattern, a different filler for 
the slot, a comparative pattern such as "too hard", and 
so forth. Although these rules use expectations about 
how the speaker might respond, they seem to have little 
to do with the expected discourse goals of the speaker; 
as a result, they would be unable to differentiate frag- 
ments that merely request information from fragments 
whose intent is to seek clarification or express surprise. 
3 SCOPE OF THE RESEARCH 
This paper describes a pragmatics-based approach for 
understanding intersentential elliptical fragments that 
occur in information-seeking dialogs. We have studied 
dialogs in which the information-seeker is attempting to 
construct a plan for accomplishing a task, but the plan 
itself will be executed at some time in the future. 
Examples of such tasks include expanding a company's 
product line, purchasing a home, pursuing a degree, or 
even taking a vacation. We have selected this class of 
dialogs because they are typical of a large percentage of 
the interactions with database management systems, 
decision support systems, and expert systems. How- 
ever, the principles presented in this paper should be 
extendible to other kinds of dialogs. 
Our hypotheses and strategies were motivated by an 
analysis of naturally occurring dialogs from several 
domains, including a radio call-in show 4 providing ad- 
vice on investments, interactions with the REL (Thomp- 
son 1980) natural language interface to a ship data base, 
and student advisement sessions. We are assuming that 
the natural language system plays the role of informa- 
tion-provider and that communication is via a typical 
terminal. Since our primary interest in this research is 
the affect of expectations and inferred knowledge on 
77 
Sandra Carberry A Pragmatics-Based Approach to Ellipsis Resolution 
ellipsis understanding, we have not considered the 
contribution of clue words (words and phrases, such as 
"Now" and "As I was saying", that contain informa- 
tion about the structure of a discourse (Reichman 1978)) 
to the processing of ellipsis. However, they certainly 
could be incorporated into our strategy. 
In this research, we are attempting to ascertain the 
extent to which pragmatic knowledge can be used to 
understand ellipsis and the limitations of such an ap- 
proach. However, we do not maintain that a natural 
language system should use only pragmatic knowledge; 
a robust system will need to coordinate syntactic, 
semantic, and pragmatic techniques in order to fully 
understand the wide variety of elliptical utterances 
employed in human communication. 
In the remainder of this paper, we argue for the use of 
pragmatic knowledge in ellipsis understanding, and 
present a pragmatics-based approach for understanding 
intersentential elliptical fragments uttered by a person 
seeking information in order to construct a plan for 
accomplishing a task. Section 4 discusses the knowl- 
edge needed to understand elliptical fragments. Section 
5 describes our overall interpretation framework, which 
is presented in detail in sections 6-8. Our strategy views 
elliptical fragments as highlighting, or bringing to atten- 
tion, aspects of the information-seeker's underlying 
task-related plan, with discourse expectations used to 
guide interpretation. Section 6 presents the discourse or 
communicative goals that we have found pursued via 
elliptical fragments in information-seeking dialogs. Sec- 
tion 7 discusses how discourse expectations develop 
during a dialog, describes how these expectations sug- 
gest discourse goals that an information-seeker might 
pursue, and presents discourse goal rules for recogniz- 
ing elliptical fragments that accomplish a particular 
discourse goal. The discourse goal rules guide how the 
fragment's highlighting of a component of the task- 
related plan should be understood; Section 8 describes 
a strategy for identifying the highlighted portion of the 
plan. Section 9 illustrates our framework with several 
examples, and sections 10 and 11 discuss future exten- 
sions to this work and our conclusions. Throughout this 
paper, the information seeker and information provider 
will be referred to as IS and IP respectively. 
4 REQUISITE KNOWLEDGE 
A speaker can felicitously employ intersentential ellip- 
sis only if he believes his utterance will be properly 
understood and his intent recognized by the listener. 
The motivation for this work is the hypothesis that 
speaker and hearer mutually believe that certain knowl- 
edge has been acquired during the course of the dialog 
and that this factual knowledge along with other proc- 
essing knowledge will be used to deduce the speaker's 
intentions. We claim that the requisite factual knowl- 
edge includes the speaker's inferred task-related plan, 
the speaker's inferred beliefs, and the anticipated dis- 
course goals; of the speaker; we claim that the requisite 
processing knowledge includes plan recognition strate- 
gies and focusing techniques. 
4.1 TASK-RELATED PLAN 
According to Grice (1957,1969), a listener must infer a 
speaker's intent in making an utterance. Now consider 
an incomplete utterance. The listener, in attempting to 
deduce the speaker's intent, will be guided by Grice's 
conversational maxims (Grice 1975). In particular, 
Grice's maxims of manner and relation suggest that the 
speaker believes his utterance is an adequate vehicle for 
conveying his intentions and that the utterance is rele- 
vant to the current dialog. A cooperative listener will 
have assimilated the preceding dialog, inferred much of 
the underlying task-related plan motivating the speak- 
er's queries, and be focused on that aspect of the task 
on which the information-seeker's attention was cen- 
tered (Grosz 1977; Carberry 1983; Carberry 1988) im- 
mediately prior to the new utterance. Given an incom- 
plete utterance, the listener can use this acquired 
knowledge to attempt to deduce the speaker's inten- 
tions by trying to fit the utterance into the partially 
constructed plan, and thereby enable the dialog to 
continue without interruption. 
This claim that the speaker's underlying task-related 
plan is an essential component of an ellipsis interpreta- 
tion strategy is further supported by research demon- 
strating the need for considering such plans in under- 
standing other types of utterances. Grosz (1977) and 
Robinson (1981) used models of the speaker's task in 
apprentice-expert dialogs to determine the referents of 
definite noun phrases and verb phrases. Perrault and 
Allen (Perrault et al. 1980) used expectations of speaker 
goals, and inference of their plans for achieving these 
goals, in the interpretation of indirect speech acts. In 
addition, Allen (Allen et al. 1980) introduced the con- 
cept of a plan-based strategy for interpreting fragmen- 
tary utterances at the outset of a dialog in a restricted 
domain. 
4.2 SHARED BELIEFS 
Shared beliefs, beliefs which the listener believes 
speaker and listener mutually hold, are a second com- 
ponent of factual knowledge required for processing 
intersentential elliptical fragments. These shared beliefs 
either represent presumed a priori knowledge of the 
domain, such as a presumption that dialog participants 
in a university domain know that each course has a 
teacher, or beliefs derived from the dialog itself. An 
example of the latter occurs if IP tells IS that CS360 is 
a 5 credit hour course; IS may not himself believe that 
CS360 is a 5 credit hour course, but as a result of IP's 
utterance and the assumption of a cooperative dialog, 
he does believe it is mutually believed that IP believes 
this. 
Real understanding requires that we identify the 
speaker's discourse goal in making the utterance. 
78 Computational Linguistics, Volume 15, Number 2, June 1989 
Sandra Carberry A Pragmatics-Based Approach to Ellipsis Resolution 
Shared beliefs, often called mutual beliefs, form a part 
of communicated knowledge used to interpret utter- 
ances and identify discourse goals in a cooperative 
dialog. The following two dialog sequences illustrate 
how IP's beliefs about IS influence understanding. 
IS: "When does CS400 meet?" 
IP: "CS400 meets on Monday from 7 pm until 9 pm." 
IS: "Who is teaching it?" 
IP: "Dr. Brown." 
IS: "At night?" 
Most people would interpret the fragmentary utterance 
"At night?" as an expression of surprise and request for 
corroboration or explanation. Now consider the follow- 
ing dialog: 
IS: "Who is teaching CS400?" 
IP: "Dr. Brown is teaching CS400." 
IS: "At night?" 
In this case, the elliptical fragment "At night?" would 
be understood as a simple request to know whether 
CS400 is meeting at night. The reason for this difference 
in interpretation is the difference in beliefs regarding IS 
at the time the elliptical fragment is uttered. In the latter 
case, IP believes it is mutually believed that IS does not 
know IP's beliefs about the meeting time of CS400. 
Since IS cannot be surprised about something he 
doesn't know, a different intention or discourse goal is 
attributed to him. 
Allen and Perrault (Perrault et al. 1980) used mutual 
beliefs in their work on indirect speech acts and sug- 
gested their use in clarification and correction dialogs. 
Sidner (1983, 1985) models user beliefs about system 
capabilities in her work on recognizing speaker inten- 
tion in utterances. Mutual beliefs appear to be a major 
influence on how human listeners interpret and respond 
to utterances. 
4.3 DISCOURSE EXPECTATIONS 
A number of researchers (Sacks et al. 1974; Labov and 
Fanshel 1977; Mann et al. 1977) have noted regularities 
in dialog and have attempted to capture these regulari- 
ties in knowledge structures that can be used for com- 
prehending utterances. For example, Mann, Moore, 
and Levin (Mann et al. 1977) used a knowledge struc- 
ture called a dialog game to model goal-related use of 
language in joint interactions such as buying/selling and 
learning/teaching. Grosz and Sidner (Grosz et al. 1986) 
and Reichman (1984) have investigated discourse struc- 
ture and have shown that a coherent discourse can be 
segmented into units that have well-defined relation- 
ships to one another. Reichman contended further that 
the existing discourse structure established expecta- 
tions about appropriate next conversational moves. 
Our analysis of naturally occurring dialog indicates 
that such expectations about appropriate next steps in 
the dialog form a third component of factual knowledge 
that plays a major role in comprehending elliptical 
fragments. The dialog preceding an elliptical utterance 
may suggest discourse or communicative goals for the 
speaker, such as seeking clarification of a question, 
expressing surprise at a question, or answering a ques- 
tion. Our transcript analysis leads us to hypothesize that 
these suggested discourse goals become shared knowl- 
edge between speaker and hearer and that, as a result, 
the listener is on the lookout for the speaker to pursue 
these anticipated discourse goals and interprets utter- 
ances accordingly. 
Consider for example the following dialog: 
Example 4 
IP: "Do you want to take CS360?" 
IS: "Who is teaching it?" 
IP: "Dr. Brown and Dr. White." 
IS: "Yes, with Dr. White." 
In this example, IP's initial query produces a strong 
anticipation that IS will pursue the discourse goal of 
providing the requested information. Therefore subse- 
quent utterances should be processed with the expecta- 
tion that IS will eventually address this goal. IS's first 
utterance is interpreted as pursuing a discourse goal of 
seeking additional information in order to answer the 
question. IS's last utterance reverts back to addressing 
the original expectation and answers the initial query 
posed by IP. 
4.4 PROCESSING KNOWLEDGE 
Our transcript analysis indicates that plan recognition 
strategies and focusing techniques are necessary com- 
ponents of processing knowledge for interpreting in- 
tersentential ellipsis. Plan recognition strategies are 
essential in order to infer a model of the speaker's 
underlying task-related plan, shown earlier to be an 
essential component of factual knowledge, and to con- 
sider possible expansions of that plan when processing 
elliptical fragments. 
Focusing techniques (Grosz 1977; Sidner 1981; Rob- 
inson 1981; McKeown 1985; Carberry 1988) are neces- 
sary in order to identify that portion of the underlying 
plan to which a fragmentary utterance refers. Consider 
again the dialog in Example 3. The focus of attention in 
this dialog is on considering the job applicants and 
evaluating them; therefore IS is most expected to con- 
tinue with this subtask before considering other aspects 
of his overall task (Carberry 1983, 1988). As a result, 
IS's fragmentary utterance "Tom's recommendation?" 
will be understood as a request for Tom's opinion about 
the suitability of the job applicants for the job of 
programming consultant. However, if the focus of at- 
tention were instead on determining the number of 
consultants needed, then IS's fragmentary utterance 
would be understood differently. 
5 FRAMEWORK FOR PROCESSING ELLIPSIS 
If an utterance is parsed as a sentence fragment, ellipsis 
processing begins. In our approach, we assume that the 
Computational Linguistics, Volume 15, Number 2, June 1989 79 
Sandra Carberry A Pragmatics-Based Approach to Ellipsis Resolution 
processor has access to a model of the preceding dialog 
containing 1. a context tree (Carberry 1988) represent- 
ing IS's inferred underlying task-related plan; 2. a stack 
containing the discourse expectations for IS, and 3. a 
belief model representing IS's inferred beliefs. These 
comprise the requisite factual knowledge discussed in 
the preceding section. 
We claim that the discourse expectations for IS 
should be used to guide interpretation of fragmentary 
utterances. The top element of the discourse stack 
represents the most immediate discourse expectation 
for IS and suggests potential discourse goals that IS 
might be expected to pursue. If IS employs a complete 
utterance, the utterance generally gives clues about IS's 
discourse goal; in this case, the discourse expectations 
play a secondary role. However, if IS employs an 
elliptical fragment, then he must intend that the frag- 
ment be interpreted according to these expectations. IS 
believes that IP expects him to pursue certain discourse 
goals and therefore that IP will be anticipating utter- 
ances directed toward these goals. If IS does not want 
his utterance interpreted according to these expecta- 
tions, then his utterance must override them, either by 
not producing an interpretation appropriate to an ex- 
pected discourse goal or by explicitly suggesting a 
different discourse goal (via the use of clue words). 
Otherwise IS is assumed to intend that his utterance be 
interpreted according to IP's expectations. 
The following alternative utterances by IS illustrate 
this hypothesis: 
IP: "Do you want to take CS360?" 
Response 1: IS: "Are you asking me ifI want to take 
it next semester?" 
Response 2: IS: "Will it be offered next semester?" 
Response 3: IS: "Next semester?" 
In Response 1, IS's utterance indicates that it is an 
attempt to clarify the question posed by IP; in Response 
2, IS's utterance simply requests information about 
CS360 in order to formulate an answer to IP's question. 
However, in Response 3, IS's utterance is an elliptical 
fragment that conceivably could produce several differ- 
ent interpretations. It might be an attempt to clarify the 
question posed by IP, it might be a request for informa- 
tion about when CS360 is offered in order to formulate 
an answer to the posed question, or it might be an 
expression of surprise that IP would ask such a question 
(for example, if CS360 is only offered next semester and 
IS already has a very full schedule). Such elliptical 
fragments do not explicitly indicate IS's discourse goal 
in uttering the fragment and thus understanding relies 
heavily on mutual beliefs, including the discourse goals 
that IS is expected to pursue. 
Thus we claim that IS must intend that his elliptical 
fragment be interpreted according to these mutually 
believed expectations. Figure 1 presents our algorithm 
for ellipsis resolution. It uses the stack of discourse 
expectations to control interpretation of elliptical frag- 
80 
ments (Loop A of Figure 1). Associated with each 
expectation is a discourse expectation rule that suggests 
discourse goals that IS might pursue (steps 2 and 3 in 
Figure 1). Each suggested discourse goal has one or 
more associated discourse goal rules that are applied in 
an attempt 1:o produce a coherent interpretation of the 
elliptical fragment (Loop B in Figure 1). We view IS's 
elliptical fragment as highlighting, or bringing to atten- 
tion, some aspect of the underlying task-related plan. 
The condition part of a discourse goal rule uses the 
aspect of the plan highlighted by the elliptical fragment, 
as identified by a plan analyzer (step 9 in Figure I), 
along with other factual and processing knowledge to 
determine whether its associated discourse goal should 
be recognized as the discourse goal being pursued by 
IS. If the condition part of a discourse goal rule is 
satisfied, the rule then produces an interpretation of the 
elliptical fragment relevant to the recognized discourse 
goal (steps 11-13 in Figure 1). The individual steps of 
the algorithm in Figure 1 will be described in the 
following sections. 
6 DISCOURSE GOALS 
We define a discourse goal intuitively as what the 
speaker is trying to do in making an utterance. The 
discourse goal itself is content independent, although its 
realization in an utterance will involve some term or 
proposition from the speaker's underlying task-related 
plan. For example, the discourse goal Express-Sur- 
prise-Obtain-Corroboration represents a speaker's at- 
tempt to convey to his listener that he is surprised at 
some proposition P and would like justification of it. 
A number of researchers have contended that a 
coherent discourse consists of segments that are related 
to one another. The structuring relations have been 
given many names, including rhetorical predicates 
(Grimes 1975, McKeown 1985), coherence relations 
(Hobbs 1979), discourse segment purposes (Grosz 
1986), and conversational moves (Reichman 1984). Our 
concept of discourse goals is closest to Grosz and 
Sidner's discourse segment purposes and Reichman's 
conversational moves. Grosz and Sidner define a dis- 
course segment purpose as the intended role of a 
discourse segment in achieving an overall discourse 
purpose. They contend that the discourse segment 
succe,~sfully serves its intended function only if the 
discourse segment purpose is recognized by the lis- 
tener. Similarly, we contend that understanding an 
elliptical fragment requires that the listener identify the 
discourse goal being pursued by the speaker in uttering 
the fragment. Reichman differentiated between utter- 
ances that continued the current discourse segment and 
utterances that constituted conversational moves to a 
segment playing a different role in the overall dialog. 
Some of her conversational moves, such as challenging 
a claim, had the flavor of goals that one might pursue in 
a dialog. Similar to our work, Reichman contended that 
Computational Linguistics, Volume 15, Number 2, June 1989 
Sandra Carberry A Pragmatics-Based Approach to Ellipsis Resolution 
L°a°P 
Input 
DISC-STACK: discourse stack containing expectations about the user's discourse 
behavior, along with the semantic representation of the utterance 
that prompted each expectation. 
SEM-REP: semantic representation of the elliptical fragment. 
CONTEXT-TREE: representation of the system's beliefs about the user's 
underlying task-related plan. 
BELIEF-MODEL: representation of the system's beliefs about the user's beliefs. 
Procedure 
I.SUCCESS ~-- FALSE; 
--Until SUCCESS = TRUE, Repeat 
2.DISC-EXP-RULE ~-- discourse expectation rule associated with Top(DISC-STACK); 
3.SUGGESTED-DISC-GOAL-SET ~--goals suggested by DISC-EXP-RULE; 
--While SUCCESS = FALSE and SUGGESTED-DISC-GOAL-SET:~ 0 Do 
4.SUGGESTED-DISC-GOAL ~ first element of SUGGESTED-DISC-GOAL-SET; 
5.SUGGESTED-DISC-GOAL-SET ~-- SUGGESTED-DISC-GOAL-SET - {SUGGESTED DISC-GOAL}; 
6.DISC-GOAL-RULE-SET ~ discourse goal rules associated with SUGGESTED DISC-GOAL; 
Loop --While SUCCESS = FALSE and DISC-GOAL-RULE-SET:/: 0 Do 
B - 7.DISC-GOAL-RULE ~ first element of DISC-GOAL-RULE-SET; 
8.DISC-GOAL-RULE-SET ~-- DISC-GOAL-RULE-SET - {DISC-GOAL-RULE}; 
9.Call plan analyzer to identify elements of the user's plan highlighted by the elliptical fragment. 
Input to Plan Analyzer 
CONTEXT-TREE: representation of system's beliefs about user's plan. 
Loop SEM-REP: semantic representation of the elliptical fragment. 
C- BELIEF-MODEL: representation of system's beliefs about user's beliefs. 
Output from Plan Analyzer 
T: term in user's plan highlighted by elliptical fragment, if any. 
P: proposition in user's plan highlighted by elliptical fragment, if any. 
CONTEXT-PROPS: conjunction of propositions forming context for T or P in the user's plan. 
10.If the conditions in DISC-GOAL-RULE are satisfied, Then 
I I.SUCCESS ~-- TRUE; 
12.DISCOURSE-GOAL ~-- SUGGESTED-DISC-GOAL; 
13.INTERPRETATION ~--interpretation generated by DISC-GOAL-RULE; 
End If-Then; 
--End While-Do; 
--End While-Do; 
14.If SUCCESS = FALSE, Then pop the top element of DISC-STACK; 
--End Until-Repeat; 
Output 
DISCOURSE GOAL: user's identified discourse goal 
INTERPRETATION: Interpretation of the elliptical fragment 
Figure 1. Algorithm for Interpreting Elliptical Fragments. 
conversational moves establish expectations about sub- 
sequent conversational moves. However, both dis- 
course segment purposes and conversational moves 
differ from our discourse goals in that a discourse 
segment may contain more than one utterance--ie., 
discourse segment purposes and conversational moves 
relate discourse segments, not utterances. In addition, 
neither Grosz and Sidner nor Reichman provide the 
details of a computational mechanism for identifying the 
role of an utterance in a dialog. Our theory does this for 
elliptical fragments in an information-seeking dialog. 
Given the discourse goal expectations, it provides a 
mechanism for understanding an elliptical fragment by 
recognizing the intentions it communicates. 
We have analyzed dialogs from several different 
Computational Linguistics, Volume 15, Number 2, June 1989 
domains and have identified 15 discourse goals that 
occur during information seeking dialogs and which 
may be accomplished via elliptical fragments. The fol- 
lowing are some of these discourse goals, with illustra- 
tive examples. 
1. Provide-For-Assimilation: IS provides information 
pertinent to formulation of his underlying task- 
related plan. 
IS: "I want to get a degree." 
"CS major." 
2. Obtain-Information: IS requests information rele- 
vant to constructing the underlying task-related 
plan or relevant to formulating an answer to a 
question posed by IP. 
81 
Sandra Carberry A Pragmatics-Based Approach to Ellipsis Resolution 
IS: "Is CS360 being offered this fall?" 
IP: "Yes." 
IS: "The instructor?" 
3. Express-Surprise-Obtain-Corroboration: IS ex- 
presses surprise regarding some proposition P and 
requests elaboration on, and justification of, it. 
IS: "What time does CS360 meet this fall?" 
IP: "Monday, Wednesday, and Friday at 8 am." 
IS: "Who is teaching it?" 
IP: "Dr. Smith." 
IS: "At 8 am?" 
4. Seek-Identify: IS is unable to satisfactorily identify 
the referent of an item in IP's utterance and 
requests help from IP in doing so. 
IS: "What is Dr. Smith teaching this fall?" 
IP: "CS360." 
IS: "The course in architecture?" 
5. Seek-Clarify-Question: IS requests information 
relevant to clarifying a question posed by IP. 
IP: "Do you want to take CS105?" 
IS: "Next semester?" 
6. Suggest-Answer-Own-Question: IS suggests an an- 
swer to his own question. IS's intent is to suggest 
that IP give particular consideration to IS's pro- 
posal in formulating an answer to the question. 
IS: "What course should I take during winter 
session?" 
"CS370?" 
7. Answer-Question-With-Restrictions: IS answers a 
Yes-No question, providing restrictions on the 
relevant underlying task-related plan. 
IP: "Would you like to take CS360?" 
IS: "At night with Dr. White." 
Although they are not explicitly accomplished via ellip- 
tical fragments, two other discourse goals play a major 
role in understanding ellipsis and will be discussed 
further in the next section: 
• Accept-Response: IP has responded to IS's request 
for information; IS accepts the response. 
• Accept-Question: IP has posed a question to IS; IS 
accepts the question. 
7 DISCOURSE COMPONENT 
7.1 DISCOURSE EXPECTATIONS 
The discourse goals of the preceding section also serve 
as discourse expectations. As argued in Section 43, 
discourse expectations play a major role in the compre- 
hension of elliptical fragments. When IS makes an 
utterance, he is attempting to accomplish a discourse 
goal; this discourse goal may in turn establish expecta- 
tions about what IS will do next. For example, if IS asks 
a question, one anticipates that IS may want to expand 
on his question by further identifying an entity or by 
clarifying the question. 
Similarly, utterances made by IP also establish ex- 
pectations for IS. These discourse expectations may 
82 
often be met implicitly as well as explicitly. Consider 
what happens when IP poses a question to IS. In a 
cooperative dialog, we expect IS to answer the ques- 
tion. But before that can occur, IS must understand the 
question and accept it as relevant and valid. Normally 
dialog participants accept such questions implicitly by 
answering the question or by proceeding to seek infor- 
mation relevant to formulating an answer. However, IS 
may be unable to accept the question posed by IP 
because he does not understand it (perhaps he is unable 
to identify some of the entities mentioned in the ques- 
tion), or because he is surprised by it. This leads IS to 
pursue discourse goals such as seeking confirmation, 
seeking the identify of an entity, seeking clarification of 
the posed question, or expressing surprise at the ques- 
tion. Thus, when IP poses a question to IS, our dis- 
course expectations are that IS will first accept the 
question (or if that is not yet possible, work towards 
accepting it) and then answer it. Thus these two expec- 
tations must be pushed onto the discourse stack and 
should serve to guide our understanding of elliptical 
fragments. 
7.2 THE DISCOURSE STACK 
The discourse stack contains expectations about IS's 
discourse behavior, along with the semantic represen- 
tation of the utterance that prompted the expectation. 
We contend that understanding elliptical fragments re- 
lies on discourse expectations that are pushed onto or 
popped from the stack as a result of utterances made by 
IS and IP. Precisely how this is accomplished in all 
cases is unclear. However, in this section, we will 
motivate the contents of the discourse stack by giving a 
set of stack processing rules that hold for simple 
utterances 5. These rules presume that the discourse 
goal accomplished by an utterance can be identified. 
Although our theory for ellipsis processing includes 
recognition of the discourse goal fulfilled by an elliptical 
fragment, we have not investigated the recognition of 
discourse goals pursued via complete sentences. This is 
an issue that requires further study and must take into 
account indirect speech acts (Perrault et al. 1980). The 
discourse stack accessed by our ellipsis processor 
(steps 2 and 14 in Figure l) represents the system's 
current beliefs about discourse expectations, but our 
implementation does not include actually constructing it 
from the dialog preceding the elliptical fragment. Our 
intent in this section is to motivate the contents of the 
discourse stack immediately preceding an elliptical frag- 
ment. 
For example, when IP asks a question of IS, IS is 
first expected to accept the question, either explicitly or 
implicitly, and then answer the question. Similarly, 
when IP answers a question posed by IS, IS is expected 
to accept the response, either explicitly or implicitly, 
and then pursue other discourse goals. This reasoning 
leads to the stack processing (SP) rules. 
Computational Linguistics, Volume 15, Number 2, June 1989 
Sandra Carberry A Pragmatics-Based Approach to Ellipsis Resolution 
SPI: If IP asks a question of IS with the discourse 
goal of seeking information, Answer-Question and 
Accept Question are pushed onto the discourse 
stack. 
SP2: When IP answers a question posed by IS, 
Accept-Response is pushed onto the discourse 
stack. 
When IS pursues a discourse goal, such as seeking 
information or clarification, one's expectation is that IS 
will continue pursuing this discourse goal with subse- 
quent utterances. This leads to the stack processing rule 
SP3: When IS actively pursues a discourse goal, the 
discourse goal is pushed onto the discourse stack. 
Although the strongest expectations are that IS will 
pursue a goal suggested by the top element of the 
discourse stack, this expectation can be passed over, at 
which point it no longer suggests expectations for 
utterances. This produces the stack popping (SPP) rule 
SP4: When IS's utterance does not pursue a goal 
suggested by the top entry on the discourse stack, 
this entry is popped from the stack. 
Other stack processing rules are similarly motivated and 
formulated. IflP asks or answers a question, SP1 or SP2 
respectively apply. If IS makes an utterance, then rule 
SP4 is applied until it fails, at which point rule SP3 is 
applied. 
7.3 SUGGESTED DISCOURSE GOALS 
We have argued that an ongoing dialog establishes 
certain discourse expectations for IS and have repre- 
sented these expectations as a discourse stack. We have 
further claimed that elliptical fragments do not explicitly 
indicate the discourse goal being pursued, and therefore 
IS must intend that elliptical fragments be interpreted 
according to mutually believed expectations. From our 
analysis of naturally occurring dialogs, we have formu- 
lated discourse expectation rules indicating how dis- 
course expectations suggest the identity of IS's dis- 
course goal. Associated with each discourse 
expectation is a rule that suggests a set of one or more 
discourse goals that IS might pursue and the order in 
which they should be considered. 
Suppose that IP has posed a question to IS. Then the 
strongest discourse expectation is that IS will under- 
stand and accept the posed question or work toward 
understanding and accepting it. In order to understand a 
question, the question must have been satisfactorily 
transmitted. If not, IS will attempt to confirm those 
components he believes may have been miscommuni- 
cated; thus the first suggested discourse goal for IS 
under this discourse expectation is Seek-Confirm. Once 
IS believes the question has been properly transmitted, 
he must be able to satisfactorily identify the referents of 
the entities mentioned in the question. If he cannot do 
so, he may attempt to obtain further identification of an 
entity; thus the second suggested discourse goal for IS 
under this discourse expectation is Seek-Identify. Once 
the components of the question are understood, IS must 
believe he fully comprehends what is being asked. If he 
does not, he must obtain clarification of the question 
before he can answer it; thus the third suggested dis- 
course goal for IS under the discourse expectation of 
accepting the question is Seek-Clarify-Question. Once 
the question is understood, IS may refuse to accept it by 
expressing surprise at an inference drawn from the 
question or at the fact that the question was asked at all. 
Thus the fourth suggested discourse goal under the 
discourse expectation of accepting the question is Ex- 
press-Surprise-Question. This leads to the following 
discourse expectation (DE) rule: 
Rule DEI: The discourse expectation Accept-Ques- 
tion suggests the following ordered set of discourse 
goals for IS: 
1. Seek-Confirm 
2. Seek-Identify 
3. Seek-Clarify-Question 
4. Express-Surprise-Question 
In a cooperative dialog, once IS has understood and 
accepted a question, IP expects him to answer it and is 
therefore on the lookout for an answer. Therefore first 
preference is given to interpretations that accomplish 
this goal; thus the first discourse goals suggested for IS 
in this situation are those that answer the question, 
directly or indirectly. If IS is unsure of the best answer 
to the question, he may suggest one or more possible 
answers and request IP's help in evaluating them; thus 
the second discourse goal suggested for IS is Suggest- 
Answer-Question. In a cooperative dialog, we expect IS 
to answer questions posed to him. If he does not do so, 
he must not have sufficient knowledge to formulate a 
good answer; in this case, we expect IS to work toward 
being able to answer the question by gathering whatever 
extra information he needs. Thus the third discourse 
goal suggested for IS under the discourse expectation of 
answering the question is Obtain-Information. This 
analysis leads to the following discourse expectation 
rule: 
Rule DE2: The discourse expectation Answer Ques- 
tion suggests the following partially ordered set of 
discourse goals for IS: 
1. Answer-Question 
Answer-Question-With-Restrictions 
Answer-Question-Suggest-Alternative 
2. Suggest-Answer-Question 
3. Obtain-Information 
If IS has been informing IP about certain aspects of his 
underlying task-related plan, our strongest expectation 
is that he will elaborate on the plan by providing further 
detail. However, in an information seeking dialog, we 
anticipate that IS is providing this knowledge as back- 
ground information in order to help IP formulate coOp- 
erative responses to subsequent questions. Thus our 
second expectation is that IS will request information 
Computational Linguistics, Volume 15, Number 2, June 1989 83 
Sandra Carberry A Pragmatics-Based Approach to Ellipsis Resolution 
about the domain in order to achieve his objective of 
constructing a plan for his underlying task. This leads to 
the following discourse expectation rule: 
Rule DE3: The discourse expectation Provide-For- 
Assimilation suggests the following ordered set of 
discourse goals for IS: 
1. Provide-For-Assimilation 
2. Obtain-Information 
Other discourse expectation rules are similarly moti- 
vated and formulated. 
As described earlier, discourse expectations are rep- 
resented in the discourse stack. These discourse expec- 
tations suggest anticipated behavior for IS and are 
analyzed in turn until a coherent interpretation for the 
elliptical fragment is identified (Loop A in Figure 1). 
The top element of the discourse stack contains the 
strongest discourse expectation at the current point in 
the dialog. The rule associated with this discourse 
expectation is used to suggest discourse goals that IS 
might pursue (steps 2 and 3 in Figure 1). Each of these 
suggested discourse goals is tried in turn (Loop B in 
Figure 1). If it is determined that IS is not pursuing any 
of the discourse goals suggested by a discourse expec- 
tation, then that discourse expectation is popped from 
the discourse stack (step 14 in Figure 1) and the process 
is repeated (Loop A in Figure 1) using the rule associ- 
ated with the new discourse expectation now residing 
on top of the stack to suggest discourse goals for IS. 
Thus, for example, when IP asks IS a question, the 
discourse expectations Answer-Question and Accept- 
Question are successively pushed onto the discourse 
stack. If IS then uses an elliptical fragment, the dis- 
course expectation Accept-Question will first be u~ed to 
suggest discourse goals that IS might be pursuing. 
However, if an interpretation of the fragment accom- 
plishing one of these suggested discourse goals cannot 
be constructed, then this discourse expectation will be 
popped from the discourse stack; this indicates that IS 
has understood and implicitly accepted the question 
asked by IP. The top element of the stack will now be 
Answer-Question and it will be used to suggest dis- 
course goals that IS might be pursuing via the fragment. 
(Note that although discourse expectations can be 
popped from the discourse stack, a cooperative dialog 
participant will never pass over the discourse goal of 
answering a question. In this research, we have as- 
sumed that the participants are always cooperative.) 
7.4 DISCOURSE GOAL RULES 
We have claimed that processing of elliptical fragments 
should be controlled by the discourse expectations in 
effect at that point in the dialog. Discourse expectations 
suggest discourse goals that IS might be expected to 
pursue. If an elliptical fragment can be interpreted as 
pursuing a suggested discourse goal, then this interpre- 
tation of the fragment should be regarded as the one 
intended by IS; otherwise IS would have overridden 
84 
these expectations, since he is aware that they will be 
used in unde, rstanding his utterances. 
Associated with each discourse goal is one or more 
discourse goal rules (step 6 in Figure 1): each rule 
applies factual and processing knowledge to analyze a 
fragment and determine whether it can be understood as 
pursuing the discourse goal with which the rule is 
associated (Loop C in Figure 1). We view fragments as 
highlighting terms or propositions in the underlying 
task-related plan; recognition of the highlighted aspect 
is described in the next section. The discourse goal rules 
determine whether IP believes that it is mutually be- 
lieved that IS might be pursuing the discourse goal 
under consideration and whether the highlighted aspect 
of the plan provides sufficient information for recogniz- 
ing the suggested discourse goal (step 10 in Figure 1). If 
so, then the elliptical fragment should be understood as 
directed toward accomplishing this suggested discourse 
goal and a related interpretation produced (steps 11-13 
in Figure 1). 
In order to demonstrate how discourse goal rules 
facilitate recognition of the discourse goal that IS is 
executing, let us examine how IS might seek identifica- 
tion of an entity appearing in IP's utterance. Figure 2 
Seek-Identify- I(IS, entityl) 
Applicability 7 Know(IS, Referent(entityl)) 
Conditions: Want(IS, Know(IS, Referent(entityl))) 
Body: Request(IS, IP, Inform(IP, IS, 
Referent (entityl))) 
Effect: Seek-Identify(IS, entity 1) 
Seek-Identify-2(IS, entityl, entity2) 
Applicability 7 Know(IS, Referent(entityl)) 
Conditions: Want(IS, Know(IS, Referent(entityl))) 
Know(IS, Referent(entity2)) 
7 Believe(IS, 7 Same-Referents 
(entityl, entity2)) 
Body: Request(IS, IP, Inform(IP, IS, 
Same-Referents(entity l, entity2))) 
Effect: Seek-Identify(IS, entityl) 
Figure 2 Two Plans for Seeking Identification of an Entity. 
presents two plans, Seek-Identify-1 and Seek-Identify- 
2, whose primary effect is the same, namely to seek 
identification of an entity. These two plans differ in that 
in the body of the first plan, IS directly requests further 
identification of the entity in question, whereas in the 
second plan, IS attempts to determine the entity by 
asking if it is the same as an entity with which he is 
familiar. Now let us consider how an elliptical fragment 
might be recognized as pursuing one of these plans. 
Suppose we have the following dialog: 
IS: ~'What CS courses are being offered during the 
summer?" 
IP: "'CS105 and CS461." 
Computational Linguistics, Volume 15, Number 2, June 1989 
Sandra Carberry A Pragmatics-Based Approach to Ellipsis Resolution 
IS: "CS461 ?" 
Suppose further that Seek-Identify has been suggested 
as a discourse goal for IS. If IS intends IP to recognize 
Seek-Identify-1 as the discourse plan that he is pursu- 
ing, then he must indicate entityl, the entity whose 
identification he is requesting. The elliptical fragment 
"CS461?" may be understood as doing this. The appli- 
cability conditions for Seek-Identify-1 specify con- 
straints that must be satisfied in order for IS to pursue 
this plan; only if IP believes that it is mutually believed 
that these constraints are satisfied can IP infer that IS 
intends him to recognize that he is pursuing this dis- 
course goal. In the above dialog, IP believes that it is 
mutually believed that IS wants to know the referent of 
CS461 since CS461 was mentioned in IP's previous 
response; it is also reasonable for IP to believe that IS 
may not know the referent of CS461. 
Let us examine this latter belief in greater detail. One 
might ask why IP did not provide an extended descrip- 
tion of CS461 in her previous response, if she believes 
that IS cannot satisfactorily identify the course from its 
department and number. This is not an issue with which 
this paper is concerned, but we will suggest one possi- 
bility. In constructing a description, IP uses knowledge 
about the listener (Appelt 1985, Goodman 1986) to 
produce a concise reference that she believes will be 
acceptable. Sometimes she will be quite certain that the 
listener will be able to identify the referent of her 
description--for example, perhaps the same description 
has been used successfully before. At other times, IP 
will be uncertain whether the description is sufficiently 
detailed; she could go to the extreme and include every 
detail at her disposal, but this may confuse the listener 
and hamper identification (Goodman 1986). Therefore, 
if IP does not know that IS can identify the referent of 
an entity from her description, she must wait for IS's 
response to this description to determine its success. In 
the case of the above dialog, IS does not implicitly or 
explicitly accept the description (he could do this by 
passing over the discourse goal of Seek-Identify) and IP 
is left with the belief that her description was unsatis- 
factory. 
In contrast, consider the following variation of the 
above dialog: 
IS: "What courses is Dr. Jones teaching next fall?" 
IP: "CS440 and CS461." 
IS: "What days is CS461 offered?" 
IP: "Monday evenings." 
IS: "What courses are being offered during the 
summer?" 
IP: "CSI05 and CS461." 
IS: "CS461?" 
It is unlikely that IS's last utterance would be inter- 
preted as seeking identification of CS461. This is be- 
cause IS implicitly accepted IP's first response, imply- 
ing that he understood the description "CS461", and 
subsequently used it himself. This leads IP to believe 
that IS can satisfactorily identify the referent of the term 
"CS461". Therefore, if IS really does want to pursue 
this discourse goal, he must override IP's belief with an 
utterance such as 
"I haven't heard of CS461. What course is it?" 
This leads to the discourse goal (DG) rule: 
Rule DG-Seek-Identify-l: Check that the following 
conditions are satisfied: 
1. IS's elliptical fragment terminates in a "?" 
2. The fragment highlights a component of IS's 
underlying task-related plan and matches a de- 
scription used in the utterance by IP that is 
closest to the top of the discourse stack. 
3. It is mutually believed that IS might not know 
the referent of this description. 
If these conditions are satisfied, then interpret the 
fragment as seeking further identification of the high- 
lighted plan component. 
The first condition in the above rule checks that IS is 
making a request (the body of the plan for Seek- 
Identify-1 in Figure 2), the second condition checks that 
IS has provided a component of his plan (which might 
be the entity whose referent is the object of the request 
in the Seek-Identify-1 plan in Figure 2) and that it is 
reasonable to believe that he wants to know its referent 
(the second applicability condition in the Seek-Identify- 
1 plan in Figure 2), and the third condition checks that 
IS does not already know what this referent is (the first 
applicability condition in the Seek-Identify-1 plan in 
Figure 2). A discourse goal rule is only invoked when 
discourse expectations suggest that IS might be pursu- 
ing its associated discourse goal. Therefore, if the 
conditions of an invoked rule are satisfied, then IS has 
provided all the information necessary for IP to believe 
that IS is pursuing that goal and IP is justified in 
interpreting the elliptical fragment accordingly. 
Similarly, if IS intends IP to recognize Seek-Identify- 
2 as the discourse plan that he is pursuing, then its 
applicability conditions must be satisfied and IS must 
indicate both entityl (the entity whose identification he 
is requesting) and entity2 (the entity with which he is 
familiar). Entity2 must be provided explicitly by the 
elliptical fragment, since it is new information not 
currently focused on in the dialog, and entityl implicitly 
by the plan element with which the fragment is associ- 
ated. This leads to the following discourse goal rule: 
Rule DG-Seek-Identify-2: Check that the following 
conditions are satisfied: 
I. IS's elliptical fragment terminates in a "?" 
2. The fragment highlights a component of IS's 
underlying task-related plan and this compo- 
nent is referenced by a description D in the 
utterance by IP that is closest to the top of the 
discourse stack. 
Computational Linguistics, Volume 15, Number 2, June 1989 85 
Sandra Carberry A Pragmatics-Based Approach to Ellipsis Resolution 
3. It is mutually believed that IS might not know 
the referent of the description D. 
4. It is mutually believed that IS does not believe 
that the fragment and the description D refer to 
different entities 6. 
If these conditions are satisfied, then interpret the 
fragment as seeking further identification of the high- 
lighted plan component. In particular, IS's fragment 
describes an entity with which he is familiar and asks 
whether this entity and the highlighted plan compo- 
nent are the same. 
One might suggest using a general plan recognition 
strategy to recognize discourse goals as well as domain 
goals. However, such a strategy has not been ade- 
quately worked out even for complete utterances, al- 
though Litman's work (Litman and Allen 1987) is a 
major step in this direction. Furthermore, as we argued 
earlier, elliptical fragments require special treatment. 
Our discourse goal rules are compilations of the kind of 
reasoning that must be done on discourse plans to 
ascertain if an elliptical fragment is intended to be 
recognized as pursuing a suggested discourse goal. 
The following are four other sample discourse goal 
rules. As mentioned earlier and discussed in detail in the 
next section, an elliptical fragment is viewed as high- 
lighting a term T or proposition P in IS's plan, possibly 
within the context of an additional set of propositions, 
which we will refer to in the following rules as CON- 
TEXT PROPS. For example, the fragment "With Dr. 
Smith?" might highlight the proposition 
P = Teach(SMITH,_ss:&SECTIONS), 
but the plan inferred from the preceding dialog might 
indicate that P should be interpreted within the context 
of the additional propositions 
Is Section Of(_ss:&SECTIONS, CS200) 
Section Offered(_ss:&SECTIONS, 
NEXT-SEMESTER) 
If IS's discourse goal were identified as Obtain Infor- 
mation, the resultant interpretation would be that IS 
wants to know whether Dr. Smith is teaching a section 
of CS200 next semester, not merely whether Dr. Smith 
is ever teaching a section of any course. The examples 
in Section 9 will illustrate the use of the following rules. 
Rule DG-Express-Surprise-Obtain-Corroboration-1: 
Check that the following conditions are satisfied: 
1. IS's elliptical fragment terminates in a "?" 
2. The fragment highlights a proposition P in the 
context of a conjunction of propositions CON- 
TEXT-PROPS in IS's underlying task-related 
plan. 
3. It is mutually believed that IS already knows 
IP's beliefs about the truth of the proposition 
(P/k CONTEXT-PROPS). 
If these conditions are satisfied, then interpret the 
fragment as expressing surprise at IP's response and 
seeking corroboration of it. In particular, IS is sur- 
prised at the known truth value of (P/~ CONTEXT- 
PROPS) in light of the new information provided by 
IP's response and the proposition P highlighted by 
IS's t'ragment. 
Rule DG-Obtain-Information: Check that the follow- 
ing conditions are satisfied: 
1. IS's elliptical fragment terminates with a "?" 
2. The fragment highlights a term T or proposition 
P in the context of a conjunction of propositions 
CONTEXT-PROPS in IS's underlying task-re- 
lated plan. 
3. It is mutually believed that IS does not know 
the value of the highlighted term or proposition. 
If these conditions are satisfied, then interpret the 
fragment as requesting information about the value of 
the highlighted term T or proposition P within the 
context of the propositions in CONTEXT PROPS. 
Rule DG-Provide-For-Assimilation: Check that the 
following conditions are satisfied: 
1. IS's elliptical fragment terminates with a .... 
2. The fragment highlights a term T or proposition 
P in the context of a conjunction of propositions 
CONTEXT-PROPS in IS's underlying task-re- 
lated plan. 
If these conditions are satisfied, then interpret the 
fragment as specifying that the user-specified term 
replace the term with which it is associated in the 
plan or that the proposition be satisfied as part of the 
plan being constructed. 
Rule DG-Answer-Question-With-Restrictions-I: 
Check that the following conditions are satisfied: 
1. IS's elliptical fragment terminates with a .... 
2. The last unanswered question asked by IP (the 
one at the top of the discourse stack) was a 
Yes-No question. 
3. The elliptical fragment highlights one or more 
propositions P in the context of a conjunction of 
propositions CONTEXT-PROPS in IS's under- 
lying task-related plan. 
If these conditions are satisfied, then interpret the 
fragment as answering yes, with the restriction that 
the proposition (P A CONTEXT-PROPS) be satis- 
fied in the underlying plan. 
8 PLAN ANALYSTS 
8.1 REPRESENTATION OF PLANS 
We view a plan as the means by which an agent can 
accomplish a nonprimitive, task-related goal. Plans are 
represented, using an extended STRIPS formalism 
(Fikes and Nilsson 1971; Allen et al. 1980), as structures 
containing applicability conditions, preconditions, a 
plan body, and effects. The plan body is a decomposi- 
tion of the plan's goal into a set of simpler subgoals, 
each of which can be accomplished by a primitive action 
or itself has an associated domain plan. Arguments in 
plans are either constants, represented as uppercase 
86 Computational Linguistics, Volume 15, Number 2, June 1989 
Sandra Carberry A Pragmatics-Based Approach to Ellipsis Resolution 
*Earn-Credit(IS, CS300, NEXT SEMESTER, _cr:&CREDITS) 
where 
Course-Offered(CS300, NEXT SEMESTER) 
Credits-Of(CS300, _cr:&CREDITS) 
*Earn-Credit-Section(IS, _ss: &SECTIONS) 
where 
Is-Section-Of(_ss: &SECTIONS, CS300) 
Section Offered(_ss:&SECTIONS, NEXT SEMESTER) 
*Learn-Material(IS, _ss: &SECTIONS, _syl: &SYLBI) 
where 
Is-Syllabus-Of(_ss: &SECTIONS, .syI:&SYLBI) I 
*Learn-From-Person(IS,_ss:&SECTIONS,_fac: &FACULTY) 
where 
Teaches(_fac:&FACULTY,_ss:&SECTIONS) 
I *Attend-Class(IS,_plc:&MTG-PLCS,MONDAY,.tme:&MTG-TIMES) 
where 
Is-Mtg-PIc(oss:&SECTIONS,_plc:&MTG-PLCS) 
Is-Mtg-Day(_ss:&SECTIONS,MONDAY) 
Is-Mtg-Time(_ss:&SECTIONS,_tme:&MTG-TIMES) 
Learn-From-Text(IS,_txt:&TEXTS) 
where 
Uses(_ss:&SECTIONS,.txt:&TEXTS) 
Figure 3 A Portion of an Expanded Context Model. 
strings, or typed variables, represented as lowercase 
strings preceded by the character "_" and followed by 
the characters ":&" and an uppercase string giving the 
variable's type. These plans are hierarchical, since 
many of the subgoals in a plan are nonprimitive and 
therefore have associated plans that may be substituted 
for them. Thus a plan can be expanded to any desired 
degree of detail, by repeatedly replacing subgoals with 
associated plans, which themselves contain constituent 
subgoals. 
We use a tree structure called a context model to 
represent the task-related plan inferred for IS from the 
preceding dialog. Each node in this tree represents a 
goal that IS has investigated achieving and, except for 
the root, is a descendant of a higher-level g0al whose 
associated plan contains the goal represented by the 
child node. For example, in the context model shown in 
Figure 3, IS is interested in achieving the subgoal of 
learning the material in a section of CS300 and has 
investigated learning from the instructor and learning 
from the text. Since only those subgoals that have been 
investigated as part of a plan for achieving a goal are 
included in the context model, a goal node can have 
different children in two context models derived from 
different dialogs. One node in the context model is 
marked as the current focus of attention and indicates 
that aspect of the task on which IS's attention is 
Computational Linguistics, Volume 15, Number 2, June 1989 
currently centered. The path from the root of the 
context model to the current focus of attention is called 
the active path and represents the global context, or 
sequence of progressively lower-level goals that led to 
the subgoal under consideration by IS. Carberry (1983, 
1988) describes how the context model is dynamically 
constructed from an ongoing dialog. 
8.2 ASSOCIATION OF FRAGMENTS 
We view elliptical fragments as highlighting, or calling 
attention to, a term or proposition in IS's underlying 
task-related plan. Given an elliptical fragment, the task- 
related plan inferred for IS must be analyzed to deter- 
mine, based on the current focus of attention, the 
particular aspect of the plan highlighted by IS's frag- 
ment. This paper will discuss three classes of elliptical 
fragments; a description of how other fragments are 
associated with plan elements is provided in Carberry 
(1985). 
We will call the matching of fragments with plan 
elements association. A constant fragment can only 
associate with terms whose semantic type is the same or 
a superset of the semantic type of the constant. Fur- 
thermore, each term in a plan has a limited set of valid 
instantiations. A constant associates with a term only if 
the system's beliefs indicate that IS might believe that 
87 
Sandra Carberry A Pragmatics-Based Approach to Ellipsis Resolution 
the uttered constant is one of the term's valid instanti- 
ations. For example, if a plan contains the proposition 
Starting-Date(AI-CONF, _date:&DATES) 
the elliptical fragment "February 2?" will associate 
with the date term in this proposition only if the system 
believes IS might believe that the starting date for the 
AI conference is in February. Recourse to such a belief 
model is necessary in order to allow for Yes-No ques- 
tions to which the answer is "No" and yet eliminate 
potential associations which a human listener would 
recognize as unlikely. Sidner (1981) employs a similar 
strategy in her work on anaphora resolution. A co- 
specifier proposed by the focusing rules must be con- 
firmed by an inference machine; if any contradictions 
are detected, other co-specifiers are suggested.) 
A propositional fragment can be of two types. The 
first contains a proposition whose name is the same as 
the name of a proposition in the plan domain. The 
second type is a more general propositional fragment, 
which cannot be associated with a specific plan-based 
proposition until after analyzing the relevant proposi- 
tions appearing in IS's plan. The semantic representa- 
tions of the utterances 
"Taught by Dr. Smith?" 
"With Dr. Smith?" 
would produce respectively the type 1 and type 2 
propositions 
Teaches(SMITH, _ss:&SECTIONS) 
Genpred(SMITH) 
The latter indicates that the name of the specific plan 
proposition is as yet unknown, but that one of its 
parameters must associate with the constant Smith. 
A proposition of the first type associates with a 
proposition of the same name if the parameters of the 
propositions associate. A proposition of the second type 
associates with any proposition whose parameters in- 
clude terms associating with the known parameters of 
the propositional fragment. 
The semantic representation of a term such as "The 
meeting time?" contains a variable term 
_tme: &MTG-TIMES 
Such a term associates with terms of the same semantic 
type in IS's plan. Note that the existing plan may 
contain constant instantiations in place of former vari- 
ables. A term fragment still associates with such con- 
stant terms. 
8.3 RETAINING ESTABLISHED CONTEXT 
It appears that humans retain as much of the established 
context as possible in interpreting intersentential ellip- 
sis. Carbonell (1983) demonstrated this phenomenon in 
an informal poll in which users were found to interpret 
the fragment in the following dialog as retaining the 
fixed media specification. 
88 
IS: "What is the size of the 3 largest single port fixed 
media disks?" 
IP: ...... 
IS: "Disks with two ports?" 
We have noted the same phenomenon in other domains. 
Thus, when an elliptical fragment is associated with a 
component of the task-related plan, the context estab- 
lished by the preceding dialog should be used to replace 
information deleted from this streamlined, fragmentary 
utterance. 
As mentioned earlier, we use a context tree to 
represent the underlying task-related plan inferred for 
IS at this point in the dialog. The set of nodes along the 
path from the root of the context tree to the current 
focus of attention form a stack of goals and associated 
plans, the topmost of which is the most recently con- 
sidered subgoal in the current focused plan; each of 
these goals is part of the plan associated with the goal 
immediately beneath it in this stack. These active nodes 
represent the established global context, and the prop- 
ositions in these nodes can have the effect of restricting 
the instantiation of variables in their ancestor nodes in 
the context tree. For example, in Figure, IS's most 
recently considered subgoal is attending class on Mon- 
day; this restricts the sections that he is currently 
considering taking to those which meet on Monday, 
thus limiting the possible instantiations of the variable 
_ss in the Learn From Person, Learn Material, and Earn 
Credit-Section subgoals higher up on the active path. 
These restrictions represent part of the context within 
which the elliptical fragment occurs and, unless over- 
ridden by infbrmation in the utterance itself, represent 
information that is missing from the sentence fragment 
but is understood by the speaker. Thus, if we use the 
propositions appearing along the active path, we can 
retain the restrictions on the variables in ancestor nodes 
when the focus of attention pops back to them. On the 
other hand, the nodes along the paths from the root of 
the context tree to the nodes at which fragment ele- 
ments associate with plan elements represent the new 
context derived from IS's fragmentary utterance. 
For example, consider the following dialog: 
IS: "What days is CS300 taught next semester?" 
IP: "Two sections of CS300 meet on Monday and 
one section meets on Tuesday." 
IS: "What time do the sections on Monday meet?" 
IP: "One section meets on Monday at 4 pm and 
another section meets on Monday at 7 pro" 
IS: "The texts?" 
A portion of the expansion of IS's inferred task-related 
plan is shown in Figure 3. Active nodes are marked by 
asterisks. The semantic representation of the fragment 
"The texts?" will contain the variable term _book: 
&TEXTS. This term associates with the term _txt: 
&TEXTS appearing at the node for the goal 
Learn-From-Text(IS, _txt:&TEXTS) 
Computational Linguistics, Volume 15, Number 2, June 1989 
Sandra Carberry A Pragmatics-Based Approach to Ellipsis Resolution 
where 
Uses(oss:&SECTIONS, _txt:&TEXTS) 
The propositions along the active path are 
Course-Offered(CS300, NEXT SEMESTER) 
Credits-Of(CS300, _cr:&CREDITS) 
Is-Section-Of(_ss:&SECTIONS, CS300) 
Section-Offered(_ss:&SECTIONS, 
NEXT-SEMESTER) 
Is-Syllabus-Of(_ss: &SECTIONS, _syl: &SYLBI) 
Teaches(_fac: &FACULTY, _ss:&SECTIONS) 
Is-Mtg-PIc(_ss:&SECTIONS, _plc:&MTG-PLCS) 
Is-Mtg-Day(_ss:&SECTIONS, MONDAY) 
Is-Mtg-Time(_ss: &SECTIONS, 
_tme:&MTG-TIMES) 
These propositions maintain the established context-- 
namely, that we are talking about the sections of CS300 
next semester that meet on Monday. The path from the 
root of the context tree to the node at which the 
elliptical fragment associates with a term in the plan 
produces the additional proposition Uses(.ss: 
&SECTIONS,_book:&TEXTS). Thus the fragment 
highlights the term .book:&TEXTS within the context 
of the conjunction of this proposition and the nine 
propositions appearing in the active nodes. (These 10 
propositions are referred to as CONTEXT-PROPS in 
the discourse goal rules of Section 7.4.) The semantics 
of this interpretation is that IS is drawing attention to 
the term _book:&TEXTS such that the conjunction of 
all 10 propositions is satisfied--namely, textbooks used 
in sections of CS300 that meet on Monday next semes- 
ter. 
8.4 IDENTIFYING THE INTENDED ASSOCIATION 
Within IS's task-related plan, there often are multiple 
components with which a fragment might associate. 
Therefore it is necessary to identify that aspect of the 
plan that IS intended to draw attention to by using the 
elliptical fragment. Grosz has claimed that communica- 
tion can be successful only if both speaker and listener 
are focused on the same subset of knowledge and if the 
listener recognizes any shifts in focus by the speaker. 
Grosz (1977) formulated rules for recognizing focus 
shifts in apprentice-expert task dialogs. McKeown 
(1985) expanded on focus rules proposed by Sidner 
(1981) to explain how speakers should organize their 
utterances when faced with a choice of topic. 
We have investigated information-seeking dialogs in 
which the information-seeker is attempting to construct 
a plan for a task to be executed at some time in the 
future. These dialogs are not constrained by the order of 
execution of the steps of the underlying task, as in 
apprentice-expert dialogs. However, such dialogs do 
Computational Linguistics, Volume 15, Number 2, June 1989 
exhibit structure in that participants typically move 
around in predictable ways to discuss different aspects 
of the plan being constructed. This structure has led to 
the development of focusing rules that predict how a 
speaker may shift his focus of attention within the 
underlying task-related plan (Carberry 1983, 1988). 
An elliptical fragment by itself often contains very 
little information and the fragment is likely to be misin- 
terpreted if its relationship to the dialog context (repre- 
sented, in our case, by the discourse stack and the 
context model) is not correctly identified. But since 
discrepancies in how participants view their knowledge 
may exist, speakers using elliptical fragments cannot 
rely on finely grained plan structures for correct inter- 
pretation. Thus small shifts in focus of attention in the 
context model are not significant in processing elliptical 
fragments. 
Although a shift in focus occurs every time one 
moves from a goal to a subgoal in the goal's associated 
plan, some of these moves may represent relatively 
small shifts in attention while others may be quite large. 
We employ the notion of focus domains in order to 
group together goals that appear to be at approximately 
the same level of focus when a plan is explicitly fo- 
cused. Moving from one goal to another in the same 
focus domain will be considered a smaller shift in 
attention than moving from a goal in one focus domain 
to one in a different focus domain. 
In a context model, a child node will often be in the 
same focus domain as its parent. However, if investi- 
gating how to achieve a goal represents a significant 
shift in focus from the plan that contains it, the goal is 
marked in the plan library as introducing a new focus 
domain. This means that all of the goal's children 
(representing subgoals in its plan) will be part of a 
different focus domain than the goal itself---ie., a move 
from discussing the goal to considering how one would 
go about achieving it represents a large shift in atten- 
tion. In Figure 4, the subgoals preceded by a "B" are in 
a different focus domain than those preceded by an 
"A". The goals of taking a particular section of a 
course, learning the material of a course, and learning 
from a particular teacher all reside in the same focus 
domain within the expanded plan for earning credit in a 
course (and therefore are in focus to approximately the 
same degree). The goal of going to the cashier's office to 
pay one's tuition also appears within this expanded 
plan; however, it is part of a different focus domain, 
introduced by the goal of paying tuition, since it does 
not come to mind nearly so readily when one thinks 
about taking a course. 7 The goal Pay-Tuition(IS,_cr: 
&CREDITS) is marked in Figure 4 as A/B, indicating 
that it is in focus domain A but that it introduces focus 
domain B (the goals in its associated plan are in focus 
domain B). 
The focusing rules use mutual beliefs about knowl- 
edge currently focused on by the dialog participants and 
expectations about probable shifts in focus to rank 
89 
Sandra Carberry A Pragmatics-Based Approach to Ellipsis Resolution 
I 
(A/B) Pay-Tuition(IS,_cr:&CREDITS) 
I 
(B) Pay(IS,_tut:&MONEY) 
where 
Costs(_cr 1: &CREDITS,_tut:&MONEY) 
I 
(B) At(IS,_off:&OFFICE,_t: &TIME) 
where 
Is-Office-Off_off: &OFFICE,CASHIER) 
B etween(_t: &TIME,_t 1: &TIME,_t2: &TIME) 
Opens-At(_off: &OFFICE,_t 1: &TIM E) 
Closes-At(_off:&OFFICE,_t2:&TIME) 
(A) Earn-Credit(IS, CS200, NEXT-SEMESTER, _cr:&CREDITS) 
where 
Course-Offered(CS200, NEXT-SEMESTER) 
Credits-Of(CS200, _cr:&CREDITS) 
I 
(A) Earn-Credit-Section(IS,_ss:&SECTIONS) 
where 
l\[s-Section-Of(_ss:&SECTIONS,CS200) 
Section-Offered(_ss:&SECTIONS,NEXT-SEMESTER) 
(A) Learn-Material(IS,_ss: &SECTIONS,_syI:&SYLBI) 
where 
Is-SylIabus-Of(oss:&SECTIONS,_syl:&SYLBI) 
(A) Learn-From-Person(IS,_sect:&SECTIONS,_fac: &FACULTY) 
where 
Teaches(..fac:&FACULTY,_ss:&SECTIONS) 
Figure 4 Part of a Context Model Illustrating Two Focus Domains. 
alternative associations of elliptical fragments with ele- 
ments of IS's underlying task-related plan. The current 
focus domain contains those goals that are most highly 
focused within IS's underlying task-related plan at the 
current time; therefore interpretations relevant to these 
goals are preferred. For example, consider the following 
dialog segment: 
IS: "Who is teaching CS440?" 
IP: "Dr. Smith." 
IS: "What textbook is being used?" 
IP: "There is no textbook. However, a set of papers 
and notes have been assembled into a course 
manual. ' ' 
IS: "Does the bookstore carry the course manual?" 
IP: "Yes." 
IS: "Any faculty discounts?" 
Faculty discounts might be available on both courses 
taken by faculty and books purchased at the bookstore. 
But humans interpret IS's elliptical fragment as asking 
whether there is a discount for faculty purchases of the 
course manual at the bookstore. This is explained by 
noting that the goal of purchasing a text is part of the 
current focus domain at the time the elliptical fragment 
is uttered, whereas the goal of paying for the course is 
not, and interpretations within the current focus domain 
or its descendants are preferred over other interpreta- 
tions in the overall plan. This leads to the rule: 
Rule-Current-Focus-Domain: Associations with ele- 
ments in the current focus domain, or within expan- 
sions of the plans associated with goals in the current 
focus domain, should be preferred. 
Although the majority of elliptical fragments require 
little shift in focus, fragments requiring significant focus 
shift do occur and are generally interpreted correctly by 
human dialog participants. Consider for example the 
following dialog: 
IS: "Who is teaching CS440?" 
IP: "Dr. Smith." 
IS: "What textbook is being used?" 
IP: "There is no textbook. However, a set of notes 
and papers have been assembled into a course 
manual. ' ' 
IS: "Does the bookstore carry the course manual?" 
IP: "Yes." 
IS: "3 credits?" 
The fragment "3 credits?" should be interpreted as 
asking whether CS440 is a 3 credit hour course. Pur- 
chasing course materials appears as a subtask in the 
plan for taking CS440, and interpreting IS's elliptical 
fragment requires shifting from the focus domain asso- 
ciated with this subtask to the focus domain associated 
with taking CS440. This is accounted for by the expec- 
tation that IS will eventually return to the higher level 
plans whose expansion led to the most recently consid- 
ered goal, leading to the following focus rule: 
Rule-Higher-Plans: Associations with elements in the 
focus domains of higher level plans whose expansion 
led to the current focus domain should be preferred 
over other associations with elements that do not 
appear in an expansion of a goal in the current 
focused domain. 
Other focusing rules are similarly motivated. 
Although unsignaled focus shifts may sometimes 
seem awkward, human dialog participants appear to 
recognize their intended meaning as long as the utter- 
90 Computational Linguistics, Volume 15, Number 2, June 1989 
Sandra Carberry A Pragmatics-Based Approach to Ellipsis Resolution 
ances are specific enough to avoid confusion. As an 
example, consider the following naturally occurring 
dialog that took place in early spring. 
Speaker 1: "Have you received the return receipt for 
your conference paper?" 
Speaker 2: "Yes, it came yesterday." 
Speaker 1: "August the m?" 
The fragment "August the m?" was correctly inter- 
preted by Speaker 2 as a request for the date that 
Speaker 2 would be attending the conference to which 
the paper had been submitted. 
In a few instances, an elliptical fragment may asso- 
ciate with two alternative elements of IS's underlying 
task-related plan, with the focusing heuristics rating 
both equally as the association intended by the speaker. 
In this case, the utterance will be interpreted as address- 
ing both aspects of the speaker's plan. This is discussed 
further in Carberry (1985). 
9 IMPLEMENTATION AND EXAMPLES 
A prototype system demonstrating this pragmatics- 
based framework for processing intersentential ellipsis 
has been implemented for a domain consisting of the 
courses, policies, and requirements for students at a 
university. The system is presented with a semantic 
representation of an elliptical fragment, a context tree 
representing the system's beliefs about the information 
seeker's underlying task-related plan, a belief model 
representing the system's beliefs about the information 
seeker's beliefs, and an initial discourse stack contain- 
ing the system's beliefs about the information seeker's 
expected discourse behavior. The fragment's semantic 
representation gives the structure of the fragment as 
discussed in Section 8.2 and its terminating punctua- 
tion. The context tree is inferred and constructed from 
the preceding dialog using the incremental plan recog- 
nition algorithm implemented in the TRACK system 
and described in Carberry (1988). Our belief model is 
very primitive and is only intended to facilitate illustra- 
tion of the ellipsis processor. The belief model contains 
propositions that the system believes the information- 
seeker believes are satisfied. These propositions may be 
fully or partially instantiated; in the latter case, they 
represent the belief that the information seeker believes 
the proposition can be satisfied by a constant selected 
from the designated type for the argument. So, for 
example, the system's belief that the information-seeker 
believes that the starting date of the AI conference is in 
February would be represented in the belief model by 
the proposition 
Starting Date(AI-CONF, 
_date: &FEBRUARY-DATES) 
where FEBRUARY-DATES is a subclass of DATES. 
The belief model also contains other beliefs, such as the 
objects and types that the information-seeker knows 
Computational Linguistics, Volume 15, Number 2, June 1989 
about and the information-seeker's beliefs about the 
types of constants and about type generalization. Since 
our research has not addressed issues such as how 
beliefs are acquired during a dialog (Kass et al. 1987), 
how they can be efficiently represented and updated in 
a user model (Finin et al. 1986), or how discourse goals 
are recognized from complete sentences, our implemen- 
tation does not include construction of the belief model 
or the discourse stack from the dialog preceding the 
elliptical fragment. 
The following examples illustrate our ellipsis resolu- 
tion strategy. 8 All rules referenced in the first example 
have been described in the preceding sections. The 
other examples illustrate a variety of discourse situa- 
tions but, due to space limitations, are presented in less 
detail than the first example. 
9.1 EXAMPLE 1: OBTAINING INFORMATION 
IS: "I want to register for a course for next semester. 
But I missed pre-registration. 
The cost?" 
In this example, the first two utterances establish a plan 
context of taking a course, with attention directed to the 
subtask of late registering for it, and the elliptical 
fragment should be interpreted as the cost of registering 
late. Since proper understanding of the fragment relies 
on inferred knowledge about IS's underlying task, this 
example requires a plan-based interpretation. 
IS's first two utterances provide information for the 
system to use in inferring IS's underlying task-related 
plan. An expansion of the context tree at this point in 
the dialog is shown in Figure 5. In the belief model, the 
only proposition about costs asserts that IS believes 
that the cost of any item is of type MONEY; therefore 
the system assumes that IS does not know the specific 
cost of individual items. The semantic representation of 
the elliptical fragment is the triple 
_costl:&MONEY Definite ? 
specifying that the fragment is a definite noun phrase 
whose head noun is of type MONEY, followed by a 
question mark. Based on the dialog preceding the 
elliptical fragment, the discourse stack contains the 
entry Provide-For-Assimilation (discourse stack rule 
SP3 in Section 7.2). Its associated discourse expectation 
rule, Rule DE3, suggests that the strongest expectation 
is for IS to continue providing information to the system 
and that the next strongest expectation is for IS to seek 
information in order to construct a plan for the task 
motivating his interaction with the system. The dis- 
course goal rule associated with Provide-For-Assimi- 
lation, Rule DG-Provide-For-Assimilation, fails to find 
an interpretation since the elliptical fragment terminates 
with a "?". Next, the discourse goal rule associated 
with Obtain Information, Rule DG-Obtain-Information, 
is invoked. It checks that the fragment terminates with 
91 
Sandra Carberry A Pragmatics-Based Approach to Ellipsis Resolution 
*(A)Earn-Credit(IS..crse:&COURSE, NEXT-SEMESTER, -cr:&CREDITS) 
where 
Course-Offered(_crse:&COURSE, NEXT-SE, MESTER) 
Credits-Of(_crse:&COURSE, _cr:&CREDITS) 
I *(A) Earn-Credit-Section(IS, _ss:&SECTIONS) 
where 
Is-Section-Of(_ss:&SECTIONS, _crse:&COURSE) 
Section-Offered(.ss:&SECTIONS, NEXT-SEMESTER) 
*(A/C)Register-Late(IS, _ss:&SECTIONS, NEXT-SEMESTER) 
*(C)Miss-Pre-Reg(IS,NEXT-SEMESTER) (C)Pay Fee(IS, LATE-REG, NEXT-SEMESTER) 
(C)Pay(IS, _lreg:&MONEY) 
where 
Costs(LATE-REG, NEXT-SEMESTER, _Ireg:&MONEY) 
Figure 5 A Portion of the Expanded Context Tree for Example 1. 
a period and calls the plan analysis component to 
associate the fragment with IS's inferred underlying 
task-related plan. 
In the expanded context tree shown in Figure 5, the 
parenthesized letters preceding goals indicate each 
goal's focus domain, and nodes on the active path are 
marked by asterisks. Immediately prior to IS's elliptical 
fragment, the current focused plan is the plan associated 
with the goal of registering late, the most recently 
considered aspect of this plan is missing pre-registra- 
tion, and the current focus domain contains the goals 
preceded by (C). IS's fragment associates with the term 
_Ireg:&MONEY in IS's inferred plan, as well as with 
terms elsewhere in parts of the expanded plan not 
shown in Figure 5. However, none of the other terms 
appear in the current focus domain, and therefore the 
association of the fragment with _Ireg:&MONEY is 
selected as most relevant to the current dialog context. 
The path from the root of the context model to the 
current focus of attention prior to the elliptical fragment 
contains the propositions 
Course-Offered(_crse:&COURSE, 
NEXT SEMESTER) 
Credits-Of(_crse:&COURSE, _cr:&CREDITS) 
Is-Section-Of(_ss:&SECTIONS, _crse:&COURSE) 
Section Offered(_ss:&SECTIONS, 
NEXT-SEMESTER) 
indicating that the fragment should be interpreted within 
this context; the path to the node at which the fragment 
associates with a term in the plan provides the addi- 
tional proposition 
Computational Linguistics, Volume 15, Number 2, June 1989 
Costs(LATE-REG, NEXT SEMESTER, 
_cost 1 :&MONEY) 
Since the belief model does not contain a proposition 
indicating that IS knows the value of the term _costl: 
&MONEY, Rule DG-Obtain Information interprets the 
fragment as seeking information about the charge for 
late registration in order to formulate the task-related 
plan; in particular, IS is requesting the value of the term 
_costl:&MONEY such that the conjunction of the fol- 
lowing propositions (CONTEXT PROPS) is satisfied: 
Course-Offered(_crse:&COURSE, 
NEXT-SEMESTER) 
Credits-Of(_crse: &COURSE, _cr: &CREDITS) 
Is-Section-Of(_ss:&SECTIONS, _crse:&COURSE) 
Section-Offered(_ss:&SECTIONS, 
NEXT SEMESTER) 
Costs(LATE-REG, NEXT SEMESTER, 
_cost 1 :&MONEY) 
9.2 EXAMPLE 2: EXPRESSING SURPRISE 
IS: "I want to take CS310 next semester. 
Who is teaching it?" 
IP: "Dr. Smith is teaching CS310 next semester." 
IS: "What time does it meet?" 
IP: "It meets at 8AM." 
IS: "With Dr. Smith?" 
This example illustrates an elliptical fragment that con- 
veys IS's surprise at IP's response and seeks elabora- 
tion and corroboration of it. The semantic representa- 
tion of the elliptical fragment is 
92 
Sandra Carberry A Pi-agmatics-Based Approach to Ellipsis Resolution 
Genpred(SMITH) Proposition ? 
specifying that the fragment is followed by a question 
mark and is a general propositional fragment that must 
associate with a plan proposition that has Smith as an 
argument. The context tree immediately prior to the 
elliptical fragment is shown in Figure 6, with asterisks 
• preceding nodes along the path from the root of the 
context model to the most recently considered goal. The 
belief model indicates that IS knows that Dr. Smith is 
teaching a section of CS310 next semester, since this 
was communicated and accepted prior to the fragment. 
The top two entries of the discourse stack are 
Accept-Response 
Obtain-Information 
indicating that IS is first expected to accept IP's re- 
sponse to his previous question and then proceed to 
seek further information in order to expand his under- 
lying task-related plan. 
The discourse expectation rule associated with Ac- 
cept-Response, similar to Rule DE1, suggests discourse 
goals of Seek-Confirm, Seek-Identify, and Express- 
Surprise-Obtain-Corroboration. The discourse goal 
rules associated with the first two of these discourse 
goals fail to produce an interpretation since the frag- 
ment does not associate with a term in 1S's plan that 
was a component of IP's previous utterance. Rule 
DG-Express Surprise-Obtain Corroboration-I checks 
that the fragment ends with a "?" and invokes plan 
analysis on the context tree shown in Figure 6. It finds 
that IS's elliptical fragment associates with the propo- 
sition Teaches(_fac:&FACULTY, _ss:&SECTIONS) 
producing the instantiated proposition 
P = Teaches(SMITH, _ss:&SECTIONS) 
In order to retain the existing dialog context, this 
proposition must be interpreted within the context of 
the propositions appearing in nodes preceded by aster- 
isks in Figure 6, namely 
Course-Offered(CS310, NEXT SEMESTER) 
Credits-Of(CS310, _cr:&CREDITS) 
Is-Section-Of(_ss:&SECTIONS, CS310) 
Section-Offered(_ss:&SECTIONS, 
NEXT SEMESTER) 
Is-Syllabus-Of(_ss:&SECTIONS, _syl: &SYLBI) 
Is-Mtg-PIc(_ss:&SECTIONS, _plc:&MTG-PLCS) 
Is-Mtg-Day(_ss:&SECTIONS, 
_day:&MTG-DAYS) 
Is-Mtg-Time(_ss:&SECTIONS, 
_tme:&MTG-TIMES) 
The discourse goal rules refer to the conjunction of the 
above eight propositions as CONTEXT PROPS. Since 
our belief model indicates that IS already knows the 
truth value of the proposition 
(P/~ CONTEXT-PROPS) 
-- namely, that Dr. Smith is teaching a section of CS310 
next semester, Rule DG-Express-Surprise-Obtain-Cor- 
*(A)Earn-Credit(IS, CS310, NEXT-SEMESTER, _cr:&CREDITS) 
where 
Course-Offered(CS310, NEXT-SEMESTER) 
Credits-Of(CS310, _cr:&CREDITS) 
*(A)Earn-Credit-Section(IS, _ss:&SECTIONS) 
where 
Is-Section-Of(_ss:&SECTIONS, CS310) 
Section-Offered(_ss:&SECTION S,NEXT-SEMESTER) 
I *(A)Learn-Material(IS,_ss:&SECTIONS,_syl:&SYLBI) 
where 
Is-Syllabus-Of(_ss:&SECTIONS,_syl:&SYLBI) 
I *(A)Learn-From-Person(IS,_ss: &SECTIONS,_fac:&FACULTY) 
where 
Teaches(-fac: &FACULTY,_ss: &SECTIONS) 
*(A)Attend-Class(IS,_plc:&MTG-PLCS,_day:&MTG-DAYS,_tme:&MTG-TIMES) 
where 
Is-Mtg-PIc(.ss:&SECTIONS,_plc:&MTG-PLCS) 
Is-Mtg-Day(_ss:&SECTIONS,_day:&MTG-DAYS) 
Is-Mtg-Time(_ss:&SECTIONS,_tme: &MTG-TIMES) 
Figure 6 The Context Tree for Example 2. 
Computational Linguistics, Volume 15, Number 2, June 1989 93 
Sandra Carberry A Pragrnatics-Based Approach to Ellipsis Resolution 
roboration-1 identifies the elliptical fragment as ex- 
pressing surprise at, and requesting corroboration of, 
IP's response. In particular, the system believes this 
surprise is a result of 
I. the new information presented in IP's response, 
namely that 8 am is the value of the term _tree: 
&MTG-TIMES in the proposition Is-Mtg-Time 
(_ss:&SECTIONS,_tme:&MTG-TIMES) in CON- 
TEXT-PROPS. 
2. the aspect of the plan highlighted by IS's elliptical 
fragment, namely the proposition 
P = Teaches(SMITH,_ss:&SECTIONS) 
Precisely the reason why this data surprises IS would 
require an additional inference mechanism; perhaps Dr. 
Smith is known to be a notoriously late riser in the 
morning or perhaps Dr. Smith holds a full-time job 
elsewhere and only teaches occasional evening courses 
at the university. 
9.3 EXAMPLE 3: POPPING THE DISCOURSE STACK 
IP: "Do you want to take CS105 next semester?" 
IS: "Who is teaching it?" 
IP: "Dr. Ames and Dr. Wilks." 
IS: "On Monday, Wednesday, Friday with Dr. 
Ames. ' ' 
This example illustrates a situation in which multiple 
expectations must be popped from the discourse stack 
in processing the fragment. IP's first utterance estab- 
lishes an expectation that IS will accept and answer the 
question asked by IP; therefore Answer Question and 
Accept-Question are pushed onto the discourse stack. 
As a result of IS's first utterance, the question is 
implicitly accepted, popping Accept Question from the 
discourse stack. IS's first utterance is a request for 
information in order to formulate an answer to IP's 
posed question, resulting in the expectation that IS will 
continue gathering such information; thus the expecta- 
tion Obtain-Information is pushed onto the discourse 
stack. IP's response answers IS's question and causes 
the expectation Accept-Response to be pushed onto the 
discourse stack. Thus the discourse stack passed to the 
ellipsis processor contains the three entries 
Accept-Response 
Obtain-Information 
Answer-Question 
The context tree at this point in the dialog would be 
similar to that shown in Figure 6, except that CSI05 
would replace CS310 and only the first four nodes 
would be on the active path. The semantic representa- 
tion of IS's elliptical fragment is the triple 
\[GenTimePred(MON-WED-FRI), 
GenPred(AMES)\] Proposition. 
specifying that the fragment is a conjunction of two 
propositions, followed by a period. 
The discourse expectation rule associated with the 
94 
top element of the discourse stack suggests the dis- 
course goals Seek-Confirm, Seek-Identify, and Express 
Surprise-Obtain-Corroboration. The discourse goal 
rules associated with these discourse goals require that 
the fragment terminate with a question mark in order to 
produce an interpretation. Since this is not the case, the 
expectation Accept-Response is popped from the dis- 
course stack. The new top element of the discourse 
stack is Obtain-Information. Its associated discourse 
expectation rule suggests the discourse goal of obtaining 
further information, since once IS pursues such a dis- 
course goal, he is expected to continue pursuing it with 
subsequent utterances. Once again, since the fragment 
terminates in a ".", the associated discourse goal rule, 
Rule DG-Obtain-Information, fails to produce an inter- 
pretation, and the expectation Obtain-Information is 
popped from the discourse stack. 
The discourse expectation rule associated with the 
expectation Answer-Question, Rule DE2, suggests, 
among others, the discourse goal of answering the 
question by providing restrictions on the underlying 
task-related plan. Rule DG-Answer-Question-With-Re- 
strictions-1 invokes plan analysis on the context tree 
produced by the preceding dialog and finds that the 
propositions comprising the fragment associate with the 
plan propositions 
Is-Mtg-Day(_ss:&SECTIONS, _day:&MTG-DAYS) 
Teaches(_fac: &FACULTY, _ss:&SECTIONS) 
resulting in the instantiated propositions 
Is-Mtg-Day(_ss:&SECTIONS, MON-WED-FRI) 
Teaches(AMES, _ss:&SECTIONS) 
The nodes along the path from the root of the context 
model to the current focus of attention prior to the 
fragment indicate that the fragment should be inter- 
preted within the context established by the additional 
propositions 
Course-Offered(CS 105, NEXT-SEMESTER) 
Credits-Of(CS 105, _cr:&CREDITS) 
Is-Section-Of(_ss:&SECTIONS, CSI05) 
Section-Offered(_ss:&SECTIONS, 
NEXT-SEMESTER) 
Is-Syllabus-Of(_ss:&SECTIONS, _syI:&SYLBI) 
Since the fragment terminates with a period and asso- 
ciates with plan propositions, discourse goal rule Rule 
DG-Answer-Question-With-Restrictions-I interprets the 
question as answering affirmatively, with the restriction 
that the conjunction of the propositions 
Course-Offered(CS 105, NEXT-SEMESTER) 
Credits-Of(CS 105, _cr:&CREDITS) 
Is-Section-Of(_ss:&SECTIONS, CS105) 
Section-Offered(_ss:&SECTIONS, 
NEXT-SEMESTER) 
Is- Syllabus-Of(_ss:&SECTIONS, _syI:&SYLBI) 
Teaches(AMES, _ss:&SECTIONS) 
Computational Linguistics, Volume 15, Number 2, June 1989 
Sandra Carberry A Pragmatics-Based Approach to Ellipsis Resolution 
Is-Meeting-Day(_ss: &SECTIONS, 
MON-WED-FRI) 
be satisfied in IS's task-related plan--namely that the 
section of CS105 he enrolls in be taught by Dr. Ames 
and meet on Monday, Wednesday, and Friday next 
semester. 
10 EXTENSIONS AND FUTURE WORK 
The main limitation of this pragmatics-based framework 
appears to be in handling intersentential elliptical utter- 
ances such as the following: 
IS: "Who is the teacher of CS200?" 
IP: "Dr. Herd is the teacher of CS200." 
IS: "CS263?" 
Obviously, IS's elliptical fragment requests the teacher 
of CS263. Our model cannot currently handle such 
fragments. This limitation is caused by our mechanisms 
for retaining dialog context; they assume that IS con- 
structs a plan for a task in a depth-first fashion, com- 
pleting investigation of a plan for CS200 before moving 
on to investigate a plan for CS263. Since the teacher of 
CS200 has nothing to do with the plan for taking CS263, 
the mechanisms for retaining dialog context will fail to 
identify "teacher-of-CS263" as the information re- 
quested by IS. 
One might argue that the elliptical fragment in the 
above dialog relies heavily on the syntactic representa- 
tion of the preceding utterance and thus a syntactic 
strategy is required for interpretation. This may be true. 
However, dialogs such as the above really investigate 
task-related plans in a kind of "breadth-first" fashion; 
for example, in the above dialog, IS is analyzing the 
teachers of each course under consideration first, and 
will then move to considering other attributes of the 
courses. It appears that our plan-based framework can 
be extended to handle many such dialogs, by reasoning 
about how IS is constructing his task-related plan 
(Ramshaw 1989). 
Litman (Litman et al. 1987) uses metaplans in her 
formalism and is able to handle the fragment in the 
above dialog. However, she views IS as changing the 
plan from consideration of CS200 to CS263, not as 
possibly considering several plans simultaneously. As a 
result, she cannot handle fragments such as IS's last 
utterance in the following dialog: 
IS: "Who is teaching CS105?" 
IP: "Dr. Smith and Dr. Jones are teaching sections of 
CS105." 
IS: "Who is teaching CS106?" 
IP: "Dr. Derr is teaching CS106." 
IS: "The meeting time of Dr. Smith's section?" 
It appears that successful handling of such fragments 
requires that we consider several potential plans in 
parallel and we are now considering mechanisms for 
doing so. 
This research has considered only interpretation of 
sentence fragments. But it appears that the plan ana- 
lyzer could be extended to provide information missing 
from a syntactically complete sentence. For example, 
suppose IS asked: 
"Who is teaching CS300 on Monday next semes- 
ter?" 
and then subsequently asked: 
"What textbooks will be used?" 
The plan analyzer could identify both the element of the 
user's plan that associates with the incompletely spec- 
ified proposition UsesPred(Unknown,_txt:&TEXTS) 
appearing in the semantic representation of the user's 
utterance and the context within which the association 
occurs. This would enable the system to recognize that 
IS wants to know the texts used by sections of CS300 
that meet next semester on Monday. This is similar to 
interpreting the elliptical fragment "The texts?" in 
Section 8.3. A related problem, that of understanding 
pragmatically ill-formed utterances (utterances that vi- 
olate the system's model of the world), is addressed in 
Carberry (1988). 
11 CONCLUSIONS 
This research has investigated the use of pragmatic 
knowledge in understanding intersentential ellipsis oc- 
curring during an information seeking dialog in a task 
domain. Our ellipsis resolution strategy uses many 
pragmatic knowledge sources, including the information 
seeker's inferred task-related plan, his inferred beliefs, 
his anticipated discourse goals, and focusing strategies. 
However, we do not contend that a natural language 
system should use only pragmatic knowledge; a robust 
system will need to coordinate syntactic, semantic, and 
pragmatic techniques in order to fully understand the 
wide variety of elliptical utterances employed in human 
communication. 
Real understanding necessitates identifying the 
speaker's discourse or communicative goal in uttering 
an elliptical fragment. We claim that a speaker who uses 
an elliptical fragment is relying heavily on discourse 
expectations and that such fragments are often used to 
accomplish discourse goals that must be marked if a 
complete sentence is used. Our work is the first to 
address the problem of recognizing the discourse goal 
fulfilled by an elliptical utterance. Our research shows 
that an ellipsis interpretation strategy controlled by 
discourse expectations and utilizing focusing heuristics 
facilitates recognition of the speaker's intent in uttering 
an intersentential elliptical fragment, including identifi- 
cation of his discourse goal. 
ACKNOWLEDGMENTS 
Some of this work was partially supported by a grant from the 
National Science Foundation, IST-8311400, and a subcontract from 
Bolt Beranek and Newman Inc. of a grant from the National Science 
Foundation, IST-8419162 
Computational Linguistics, Volume 15, Number 2, June 1989 95 
Sandra Carberry A Pragmatics-Based Approach to Ellipsis Resolution 
I would like to thank Ralph Weischedel for his encouragement and 
suggestions and Lance Ramshaw for many helpful discussions. I also 
appreciate the comments and suggestions of Kathy McCoy, Alan 
Pope, Ralph Weischedel, and the anonymous reviewers on drafts of 
this paper. 
NOTES 
1. Taken from Flowers and Dyer (1984). 
2. A speech act is an action performed with an utterance, such as the 
acts of informing and making a promise, although other require- 
ments may need to be met for the action to be successfully 
executed (Austin 1962). 
3. Thanks to Ralph Weischedel for this example. 
4. These transcripts were provided by the Department of Computer 
Science of the University of Pennsylvania. 
5. These rules capture the behavior exhibited in the dialogs we 
studied. However, we have no experiments to indicate whether 
the model is psychologically valid. 
6. Note that --1 Believe(IS, --7 P)4-~ Believe(IS,P) 
7. This is similar to Grosz's focus spaces and the notion of an object 
being in implicit focus. 
8. Some stack processing rules that were not discussed in Section 
7.2 would influence the discourse stack but not affect how the 
elliptical fragments in the examples are understood. They have 
been omitted to keep the presentation short and to focus attention 
on the important aspects of the ellipsis processing strategy. 
Computational Linguistics, Volume 15, Number 2, June 1989 

REFERENCES 
Allen, James F. and Perrault, C. Raymond 1980 .Analyzing Intention 
in Utterances. Artificial Intelligence 15:143-178. 
Appelt, Douglas E. 1985 Planning English Refe.rring Expressions. 
Artificial Intelligence 26(1): 1-33. 
Austin, J. L. 1962 How To Do Things With Words. Harvard Univer- 
sity Press, Cambridge, MA. 
Carberry, Sandra 1983 Tracking User Goals in an Information- 
Seeking Environment. In: Proceedings of The National Confer- 
ence on Artificial Intelligence, Washington, DC: 59-63. 
Carberry, Sandra 1985 Pragmatic Modeling in Information System 
Interfaces. Ph.D. thesis, University of Delaware, Newark, DE. 
Carberry, Sandra 1988 Modeling the User's Plans and Goals. Com- 
putational Linguistics Journal 14(3):23-27. 
Carbonell, Jaime G. 1983 Discourse Pragmatics and Ellipsis Resolu- 
tion in Task-Oriented Natural Language Interfaces. In: Proceed- 
ings of the 21st Annual Meeting of the Association for Computa- 
tional Linguistics, Boston, MA: 164-168. 
Fikes, R. E. and Nilsson, N. J. 1971 STRIPS: A New Approach to the 
Application of Theorem Proving to Problem Solving. Artificial 
Intelligence 2: 189-208. 
Finin, Tim and Drager, David 1986 GUMS: A General User Modeling 
System. In: Proceedings of the Sixth Canadian Conference on 
Artificial Intelligence, Montreal, Canada: 24-30. 
Flowers, M. and Dyer, M. E. 1984 Really Arguing with your Com- 
puter. Proceedings of the National Computer Conference; 653- 
659. 
Goodman, Bradley A. 1986 Reference Identification and Reference 
Identification Failures. Computational Linguistics; 273-305. 
Grice, H. Paul 1957 Meaning. Philosophical Review 56: 377-388. 
Grice, H. Paul 1969 Utterer's Meaning and Intentions. Philosophical 
Review 68: 147-177. 
Grice, H. Paul 1975 Logic and Conversation. In: Cole, P. and Morgan, 
J. L. (eds.), Syntax and Semantics II1: Speech Acts, Academic 
Press, New York City, NY: 41-58. 
Grimes, J. E. 1975 The Thread of Discourse. Mouton. 
Grosz, Barbara J. 1977 The Representation and Use of Focus in a 
System for Understanding Dialogs. In: Proceedings of the Inter- 
national Joint Conference on Artifical Intelligence, Cambridge, 
MA: 67-76. 
Grosz, Barbara J. 1979 Utterance and Objective: Issues in Natural 
Language Processing. In: Proceedings of the International Joint 
Conference on Artificial Intelligence, Tokyo, Japan: 1067-1076. 
Grosz, Barbara and Sidner, Candace 1986 Attention, Intention, and 
the Structure of Discourse. Computational Linguistics, 12(3): 
175-204. 
Hendrix, Gary G.; Scerdoti, Earl D.; Sagalowicz, Daniel; and 
Slocum, Jonathan 1978 Developinng a Natural Language Interface 
to Complex Data. ACM Transactions on Database Systems 3(2): 
105-147. 
Hobbs, Jerry R. 1979 Coherence and Coreference. Cognitive Science 
3: 67-90. 
Kass, Robert and Finin, Tim 1987 Rules for the Implicit Acquisition of 
Knowledge About the User. In: Proceedings of the Sixth National 
Conference on Artificial Intelligence, Seattle, WA: 295-300. 
Labov, W. and Fanshel, D. 1977 Therapeutic Discourse: Psychother- 
apy as Conversation. Academic Press, New York City, NY. 
Litman, Diane J. 1986. Understanding Plan Ellipsis. In: Proceedings 
of the Fifth National Conference on Artificial Intelligence, Phila- 
delphia, PA: 619--624. 
Litman, Diane and Allen, James 1987 A Plan Recognition Model for 
Subdialogues in Conversation. Cognitive Science 11:163-200. 
Mann, William; Morre, James; and Levin, James 1977 A Comprehen- 
sion Model for Human Dialog. In: Proceedings of the Fifth 
International Joint Conference on Artificial Intelligence, Cam- 
bridge, MA: 77-87. 
McKeown, Kathleen R. 1985 Text Generation. Cambridge University 
Press, Cambridge, England. 
Perrault, R. and Allen, J. 1980 A Plan-Based Analysis of Indirect 
Speech Acts. American Journal of Computational Linguistics, 
167-182. 
Pollack, M.; Hirschberg, J. ; and Webber, B. 1982 User Participation 
in the Reasoning Processes of Expert Systems. In: Proceedings of 
the Second National Conference on Artificial Intelligence, 
(AAAL) Pittsburgh, PA: 358-361. 
Ramshaw, Lance 1989 A Metaplan Model for Problem-Solving Dis- 
course. In: Proceedings of the Fourth Conference of the European 
Chapter of the Association for Computational Linguistics, 
Manchester, England. 
Reichman, Rachel 1978 Conversational Coherency. Cognitive Sci- 
ence 2:283-327. 
Reichman, Rachel 1984 Extended Person-Machine Interface. Artifi- 
cial Intelligence 22:157-218. 
Robinson, Ann 1981 Determining Verb Phrase Referents in Dialogs. 
American Journal of Computational Linguistics; 1-18. 
Sacks, H.; Schegloff, E. A.; and Jefferson, G. 1974 A Simplest 
Systematics for the Organization of Turntaking for Conversation. 
Language 50: 696--735. 
Sidner, Candace L. 1981 Focusing for Interpretation of Pronouns. 
American Journal of Computational Linguistics; 217-231. 
Sidner, Candace L. 1983 What the Speaker Means: The Recognition 
of Speakers' Plans in Discourse\] Computers and Mathematics 
With Applications 9(1):71-82. 
Sidner, Candace L. 1985 Plan Parsing for Intended Response Recog- 
nition in Discourse. Computational Intelligence; 1-10. 
Thompson, Bozena H. 1980 Linguistic Analysis of Natural Language 
Communication with Computers. In: Proceedings of the 8th 
International Conference on Computational Linguistics, Tokyo, 
Japan: 190-201. 
Waltz, David. 1978. An English Language Question Answering Sys- 
tem for a Large Relational Data base. Communications of the 
Association for Computing Machinery 21(7):526-539. 
Weischedel, Ralph and Sondheimer, Norman K. 1982 An Improved 
Heuristic for Ellipsis Processing. In: Proceedings of the 20th 
Annual Meeting of the Association for Computational Linguistics, 
Toronto, Ontario, Canada: 85-88. 
