Reference and Pragmatic Identification 
Douglas E. AppeIt 
Artificial Intelligence Center 
Center for the Study of Language and Information 
SRI International 
1 Identification Constraints. 
Kronfeld \[9\] has presented the outline of a theory of referring that provides an excellent 
framework for explaining how referring actions operate in various contexts. However, it 
also raises some important questions, one of which I will examine in greater detail in this 
paper. This question is "Where do identification constraints come from?" 
According to Kronfeld's model, the literal goal of a referring action is to make it 
mutually believed that "identification" of some entity is required. This idea is similar to 
that advanced by Cohen \[2,3\], who argues in favor of analyzing referring as the illocutionary 
act of requesting as opposed to a propositional act, whereas Kronfeld's model retains 
aspects of the propositional act analysis. Regardless of the details of the proposal adopted, 
it does little good to say merely that a referring action requires identification of the referent, 
because the precise facts that must be known for a hearer to say that he has identified a 
referent is different in practically every case. 
In the model under consideration, an individual is represented to an agent by an in- 
dividuating set of terms, each believed to denote the individual. The ultimate goal of a 
referring action is to induce the hearer to identify a subset of one of his individuating sets 
that satisfies a number of identification constraints. The speaker and hearer must mutually 
know what the relevant identification constraints are in the current situation. Given this 
general theory of referring, the key problem becomes the explanation of how the speaker 
and hearer can agree on what identification constraints are currently applicable. 
The point of this paper is that identification constraints come from a variety of sources, 
including knowledge about actions, general world knowledge, particular facts about the sit- 
uation at hand, the semantic content of the referring expression, and principles of discourse. 
Each of these is an important area of analysis in its own right. I shall suggest how the 
various aspects of reference addressed by the members of the TINLAP-3 reference panel fit 
together under this general framework. 
149 
2 Constraints from World Knowledge 
Goodman \[6\] states that "Reference identification is a search process where a listener 
looks for something in the world that satisfies a speaker's uttered description." One can 
argue that this definition is too restrictive because it does not apply to situations in which 
an epistemological notion of identification is inappropriate. However, for task oriented 
dialogues such as those of Goodman's protocols, it is correct. The speaker and hearer 
are cooperating on a task that involves physical manipulation of assorted parts and tools. 
A reasonable theory of action would imply that physical manipulation of objects requires 
perception of the objects by the agent, and such a theory would be mutually believed by 
all agents. Therefore, the requirement that the individuating set contain a term resulting 
from some perceptual action would apply to nearly every reference to material objects in 
this domain. 
Goodman's research is centered primarily on the problem of satisfying the referring 
request by carrying out the identification plan. Some of the bizarre referring expres- 
sions obtained from Goodman's protocols \[5\] are quite reasonable from the standpoint of 
achieving the literal referring goals. When the speaker used the referring expression "the 
champagne top sort of looking bottom" to refer to the tube base of the water pump, it 
is clear that (1) he inter/ded the hearer to perceive the part, because he was asking the 
hearer to manipulate it, and (2) the referring description, consisting of perceptual de- 
scriptors, suggests a plan of visually observing objects in the domain and comparing their 
characteristics to those indicated by the description. The hearer in Goodman's protocol 
was unable to identify the intended object given this odd description, which demonstrates 
the need for the speaker to take both the satisfaction plan as well as the literal goals into 
account when planning a referring expression. 
3 Constraints from Definiteness 
The use of a definite determiner in a referring expression introduces an additional con- 
straint on the bearer's individuating set: the individuating set must exist at the time of 
the utterance, or it must be implicitly associated \[4\] with an existing individuating set, i.e. 
its existence can be inferred from its association with an existing entity. This constraint 
prevents the hearer from creating an individuating set containing only the speaker's refer- 
ring expression, which would amount to hypothesizing an entirely new entity. There is no 
such constraint associated with the indefinite determiner, which leaves the hearer free to 
hypothesize new individuating sets in the absence of any other prohibitive constraints. 
150 
Other constraints may be brought to bear on the individuating set as well. For exam- 
ple, if the speaker is requesting the hearer to manipulate the entity introduced with the 
indefinite article, a perceptual term must be part of the individuating set. For example, if 
the speaker says "There is a philips screwdriver in the toolbox that you can use to fix the 
pump," the hearer's individuating set must contain a perceptual term denoting some exist- 
ing object that he can perceive. Contextual information can sometimes be strong enough 
to imply a very strict criterion for referent identification. For example, at a testimonial 
dinner honoring John, a speaker says "We are gathered here to honor a gentleman and a 
scholar." The hearer must know there is only one person honored at the banquet, and that 
is John. Therefore, the individuating set specified must be the same as his individuating 
set for John, except that it must contain the descriptors "gentleman" and "scholar." The 
speaker has exploited the overwhelming contextual influence to produce an expression that 
performs both informing and referring functions. This strategy is called action subsump- 
tion \[1\]. Dahl \[4\] discusses several more complex situations in which the use of an indefinite 
noun phrase is not permitted to introduce new individuals. 
4 Constraints from Discourse 
A particularly interesting set of referring expressions are those that also have anaphoric 
connection to the preceeding discourse. Not all anaphoric expressions are referring ex- 
pressions. For example, in the sentence "No AI researcher will admit that he is wrong." 
neither the pronoun he nor its antecedent is a referring expression. However, pronouns 
and anaphoric definite noun phrases are frequently referring expressions. Because pronouns 
must refer anaphorically (or to some very salient object in the context), the identification 
constraints that apply to a pronominal referring expression are simple to state: The active 
individuating set must contain a term from the individuating set intended by a previous 
reference to the same individual, with gender and number providing additional constraints 
on the possible referent. 
It is not so simple, however, to state how the saris/action of the anaphoric identification 
request takes place. Much research in re~ent years has been devoted to this problem, 
including (to mention only a few instances) recent work on discourse context and centering 
by Grosz, Joshi, Sidner, and Weinstein \[7,8\], and Webber \[10,11\]. It is far beyond the 
scope of this paper to discuss this work here, or to add anything to it. The reader should 
bear in mind that the principles of centering and the construction of discourse models, 
event/situation structures, etc. are all mutually known to the speaker and hearer in a 
dialogue. The speaker takes these principles into account when reasoning about how the 
151 
hearer can formulate a plan to identify the referent of an anaphoric referring expression. 
Identification constraints from multiple sources are necessary to explain changes that 
may take place in the identification constraints applicable to different instances of coref- 
erential expressions. Consider the following sequence of utterances: 
I am looking for a screwdriver. 
It has a green handle. 
Have you seen it recently? 
In the first sentence, no constraints apply to the identification of the referent of "a screw- 
driver." The hearer therefore constructs a new individuating set to represent it. In under- 
standing the second sentence, the hearer uses the centering algorithm to determine that 
the intended individuating set for the pronoun is the same as the one intended in the first 
sentence. The fact that the hearer intends the same individuating set for the pronoun 
in the third sentence can also be determined from the centering algorithm. However, the 
fact that the hearer is asked if he has seen the object implies that an additional identi- 
fication constraint must be imposed on the individuating set at that time: the referent 
must be perceptually identified. A cooperative speaker must reason that the hearer has 
enough knowledge to satisfy the identification request before deciding that "it" constitutes 
an appropriate referring expression. 
5 Conclusion 
If referring is to be regarded as an action that requires the hearer to pragmatically identify 
the referent of a description, then it is important to describe how it is that the speaker 
and hearer know what pragmatic identification means in a given situation. This paper 
suggests that the situation dependent meaning of identification follows from general world 
knowledge, the syntactic and semantic structure of the referring expression itself, and 
principles of discourse anaphora resolution. This is by no means an exhaustive analysis of 
the ways in which identification conditions are recognized, but is rather intended to provide 
the first steps toward the analys{s of reference in a framework that links the results several 
diverse research programs. 
Acknowledgements 
This research was supported by the National Science Foundation under grant DCR- 
8407238. The author is grateful to Amichai Kronfeld for comments on the draft of this 
152 
article, and for stimulating discussion of these and related issues. 
References 
\[1\] Douglas E. Appelt. Planning English Sentences. Cambridge University Press, Cam- 
bridge, England, 1985. 
\[2\] Philip R. Cohen. The need for identification as a planned action. In Proceedings of the 
Seventh International Joint Conference on Artificial Intelligence, pages 31-36, 1981. 
\[3\] Philip R. Cohen. Referring as requesting. In Proceedings of the Tenth International 
Conference on Computational Linguistics, pages 207-211, 1984. 
\[4\] Deborah Dahl. Determiners, entities, and contexts. In this volume. 
\[5\] Brad Goodman. Repairing reference identification failures by relaxation. In Pro- 
ceedings of the 23rd Annual Meeting, pages 204-217, Association for Computational 
Linguistics, 1985. 
\[6\] Bradly A. Goodman. Reference and reference failures. In this volume. 
\[7\] Barbara J. Grosz, A. Joshi, and S. Weinstein. Providing a unified account of defi- 
nite noun phrases in discourse. In Proceedings of the Twenty-first Annual Meeting, 
pages 44-50, Association for Computational Linguistics, 1983. 
\[8\] Barbara J. Grosz and Candace L. Sidner. Discourse structure and the proper treat- 
ment of interruptions. In Proceedings of the Ninth International Joint Conference on 
Artificial Intelligence , pages 832-839, 1985. 
\[9\] Amichai Kronfeld. Goals of referring acts. In this volume. 
\[10\] Bonnie L. Webber. Description formation and discourse model synthesis. In Proceed- 
ings of TINLAP-2, pages 42-50, 1978. 
\[11\] Bonnie L. Webber. Event reference. In this volume. 
153 
