<?xml version="1.0" standalone="yes"?>
<Paper uid="P86-1029">
  <Title>DONNELLAN'S DISTINCTION AND A COMPUTATIONAL MODEL OF REFERENCE</Title>
  <Section position="1" start_page="0" end_page="0" type="metho">
    <SectionTitle>
DONNELLAN'S DISTINCTION
AND A COMPUTATIONAL MODEL OF REFERENCE
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
  <Section position="2" start_page="0" end_page="0" type="metho">
    <SectionTitle>
SRI International and
333 Ravenswood Avenue
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
ABSTRACT
</SectionTitle>
    <Paragraph position="0"> In this paper, I describe how Donnellan's distinction between referential and attributive uses of definite descriptions should be represented in a computational model of reference. After briefly discussing the significance of Donnellan's distinction, I reinterpret it as being three-tiered, relating to object representation, referring intentions, and choice of rehrring expression. I then present a cognitive model of referring, the components of which correspond to this analysis, and discuss the interaction that takes place among those components. Finally, the implementation of this model, now in progress, is described.</Paragraph>
  </Section>
  <Section position="4" start_page="0" end_page="186" type="metho">
    <SectionTitle>
INTRODUCTION
</SectionTitle>
    <Paragraph position="0"> It is widely acknowledged that Donnellan's distinction \[7\] between referential and attributive uses of definite descriptions must be taken into account in any theory of reference. There is not yet agreement, however, as to where the distinction fits in a theoretical model of definite noun phrases. For Cohen \[4\], the intention that the hearer identify a referent constitutes a crucial difference between the referential and the attributive. Barwise and Perry \[3\], on the other hand, treat their value-loaded/value-free distinction as the central feature of the referential versus the attributive. However, as pointed out by Grosz et al. \[9\], this analysis ignores an essential aspect of Donnellan's distinction, namely, the speaker's ability, when using a description referentially, to refer to an object that is independent of the semantic denotation.</Paragraph>
    <Paragraph position="1"> The problem of determining the correct interpretation of Donnellan's distinction is of considerable importance. First, Donnellan's distinction seems to violate the principle that reference to physical objects is achieved by virtue of the descriptive content of referring expressions. This principle can be found practically everywhere -- for example, in Frege's sense and reference, Rusell's theory of descriptions, and Searle's speech acts. In the referential use of definite descriptions, however, reference seems to be established independently of descriptive content. If I say ~The man over there with a glass of white wine is..., ~ I may be successful in my act of referring -- regardless of whether the person over there is a man or a woman, the glass is full of wine or grape juice, the color of the beverage is white or red, and so on. This, if accepted, has far-reaching consequences for the meaning of referring expressions, for the logical structure of propositions, and for the theory of propositional attitudes.</Paragraph>
    <Paragraph position="2"> Second, the referential/attributive distinction forces us to reconsider the division between semantics and pragmatics. It seems that a speaker's intentions in using a referring expression do make a semantic difference. If I say ~Smith's murderer is insane,&amp;quot; meaning that whoever murdered Smith is insane (the attributive case), what I say is true if and only if the one and only murderer is insane. If, on the other hand, my intention is to use the definite description referentially (referring to, say, Tom, who is accused of being the culprit), what I say is true if and only if Torn is indeed insane -- whether he is the murderer or not. Unless we understand the interaction between conventional meaning and a speaker's intentions in such cases, we cannot hope to construct an adequate model of referring and language use in general.</Paragraph>
    <Paragraph position="3"> Finally, Donnellan's distinction brings to the fore the role of identification in the speech act of referring. Both Strawson and Searle (\[17,16\]) attempted to analyze referring in terms of identification and identifying descriptions. But Donnellan has pointed to what seems to be a clear distinction between eases in which identification is required (referential use) and thor in which it is not (attributive use). This calls for a new analysis of the speech act of referring, one that does not rely on identification as a central concept, l In this paper, I present a general framework for treating Donnellan's distinction. In particular, I contend the following: 1. The apparent simplicity of the referential/attributive distinction masks three aspects of the problem of reference.</Paragraph>
    <Paragraph position="4"> In a sense, it is not one distinction but three: the first has to do with representations of objects, the second -with referring intentions, the third -- with the choice of referring expressions.</Paragraph>
    <Paragraph position="5"> 2. These three distinctions are independent of one another, and should be handled separately. Each is relevant to a different component of a plan-based model of reference: the data base, the planner, and the utterance generator, respectively.</Paragraph>
    <Paragraph position="6"> 3. Although the three distinctions are mutually independent, tThese comments, naturally, only touch the surface. For an extensive discussion of the significance of the referential/attributive distinction see my thesis \[141. For a discussion of the role of identification in referring, see the paper coauthored by Appelt and me on this topic 12\].</Paragraph>
    <Paragraph position="7">  they of course interact with one another. The notion of a conversationally relevant description provides a basis for explaining how the interaction operates.</Paragraph>
    <Paragraph position="8"> In the following sections, the three aspects are presented, their interactions discussed, and an initial attempt to achieve an implementation that takes them into account is described.</Paragraph>
  </Section>
  <Section position="5" start_page="186" end_page="186" type="metho">
    <SectionTitle>
CRITERIA
</SectionTitle>
    <Paragraph position="0"> How is the referential to be distinguished from the attributive? Two criteria are usually offered: 1. Even though, when used attributively, the description must denote the intended referent, in the referential use this is not necessary.</Paragraph>
    <Paragraph position="1"> 2. In the referential use, the speaker has a particular object in mind, whereas in the attributive he does not.</Paragraph>
    <Paragraph position="2"> These criteria have been taken to be equivalent: any use of a definite description that is referential according to one criterion should also be classified as referential according to the other (and similarly for the attributive use). However, the equivalence of the two criteria is really an illusion: some uses of definite descriptions are referential according to one criterion, but attributive according to the other. For example, let us suppose that John, a police investigator, finds Smith's murdered body, and that there are clear fingerprints on the murder weapon. Now consider John's utterance: &amp;quot;The man whose fingerprints these are, whoever he is, is insane.&amp;quot; Note that John intended to speak of Smith's murderer, and he may very well have been successful in conveying his intended referent, whether or not the fingerprints indeed belonged to the murderer. Hence, according to the first criterion, the description, &amp;quot;The man whose fingerprints these are,&amp;quot; was used referentially. On the other hand, John did not have any particular person in mind. Hence, according to the second criterion, the description must have been used attributively.</Paragraph>
    <Paragraph position="3"> Many, including Donnellan, regard the second criterion as the more significant one. But even this criterion is given two conflicting interpretations. On the one hand, ~having a particular object in mind&amp;quot; is taken as an epistemic concept: this view holds that one can have a particular object in mind while referring only if one knows who or what the referent is. On the other hand, the criterion also receives what I call the modal interpretation. According to this reading, the referential use of a definite description is simply tantamount to employing the description as a rigid designator. Obviously, the two interpre.</Paragraph>
    <Paragraph position="4"> tations are not equivalent. As Kaplan demonstrates \[lll, one can use a description as a rigid designator without having any idea who the referent is.</Paragraph>
    <Paragraph position="5"> Thus, there are three aspects of Dounellan's distinction that should be carefully separated. These aspects can be represented in terms of three dichotomies: * Having knowledge of an object versus not having such knowledge (the epistemie distinction).</Paragraph>
    <Paragraph position="6"> ,, Using a description as a rigid designator versus using it as a nonrigid one (the modal distinction).</Paragraph>
    <Paragraph position="7"> s Using a definite description &amp;quot;the ~&amp;quot; to refer to whoever or whatever the ~ may be, versus using &amp;quot;the ~&amp;quot; to refer to an object z, whether or not z is indeed the ~ (the speech act distinction).</Paragraph>
  </Section>
  <Section position="6" start_page="186" end_page="187" type="metho">
    <SectionTitle>
THREE COMPONENTS
</SectionTitle>
    <Paragraph position="0"> The epistemic, modal, and speech act distinctions correspond to three components that a plan-based model of reference must possess, z Any such model must contain the following:  1. A database that includes representations of objects 2. A planner that constructs strategies for carrying out referring intentions 3. An utterance generator that produces referring expres null sions Let us call these the database, the planner, and the utterancegenerator, respectively. The next three sections describe a cognitive model of referring that incorporates these components.</Paragraph>
    <Section position="1" start_page="186" end_page="187" type="sub_section">
      <SectionTitle>
Object Representations
</SectionTitle>
      <Paragraph position="0"> Objects are represented to agents by terms. These terms are grouped into individuatin9 sets. An individuating set for an agent S is a maximal set of terms, all believed by S to be denoting the same object. For example, for John, the police investigator, the set {Smith'n murderer, the man who~e fingerprints these are} is an individuating set of Smith's murderer. The incredibly complex cluster of internal representations under which, for instance, John's mother would be represented to him is also an individuating set, although it would be impractical to enumerate all the terms in this set.</Paragraph>
      <Paragraph position="1"> An individuating set is grounded if it contains either a perceptual term or a term that is the value of a function whose argument is a perceptual term. For example, a set containing the description &amp;quot;your father&amp;quot; is grounded, since it contains a terms that is the result of applying the function FATHER-OF to a perceptual term representing you.</Paragraph>
      <Paragraph position="2"> It should be emphasized that an individuating set is the result of the speaker's beliefs, not a mirror of what is actually the case. A speaker may possess two distinct individuating sets that, unbeknownst to him, determine the same object (e.g., Oedipus's representations of his mother and his wife). On the other hand, a speaker may possess an individuating set containing two or more terms that actually denote different objects. Moreover, the object that an agent believes to be denoted by the terms of some individuating set may not exist in the actual world.</Paragraph>
      <Paragraph position="3"> Whether or not an agent can have knowledge of the referent, or know who or what the referent is (the epistemic distinction}, depends on the nature of the relevant individuating set. In a computational model, we can place a number of restrictions on individuating sets to reflect various epistemological intuitions. For example, we may require that, for an agent to be able to manipulate an object, the relevant individuating set must contain a perceptual term, or that, for an agent to know eFor a plan-based model of referring, definite noun phrases, and speech acts in general, see articles by Appelt, Cohen, Cohen and Levesque, Cohen and Perrault (\[1,4,.5,6\]).</Paragraph>
    </Section>
  </Section>
  <Section position="7" start_page="187" end_page="187" type="metho">
    <SectionTitle>
DISTINCTION INTERPRETATION
</SectionTitle>
    <Paragraph position="0"> who a certain person is (relative to purpose P), the relevant individuating set must include a privileged term determined by P, or that, for an agent to have knowledge o fan object, the relevant individuating set must be grounded, and so on.</Paragraph>
    <Paragraph position="1"> Since individuating sets are part of the database, this is where the epistemlc distinction belongs.</Paragraph>
    <Section position="1" start_page="187" end_page="187" type="sub_section">
      <SectionTitle>
Referring Intentions
</SectionTitle>
      <Paragraph position="0"> A speaker may have two distinct types of referring intentions.</Paragraph>
      <Paragraph position="1"> First, he may select a particular term from the relevant individuatlng set, and intend this term to be recognized by the hearer. Second, the speaker may intend to refer to the object determined by an individuating set, without intending any particular term from the set to be part of the proposition he wants to express. Consider, for example, the following two statements:  In making both statements, a speaker would normally be referring to Shakespeare. But note the difference in referring intentions between the two: in the first statement, the speaker selects a particular aspect of Shakespeare, namely, the fact that he is the author of Othello, and intends the hearer to think of Shakespeare in terms of this aspect. In the second statement, the speaker does not select any particular aspect of Shakespeare from the relevant individuating set. Indeed, he may not care at all how the hearer makes the connection between the name &amp;quot;Shakespeare&amp;quot; and the referent.</Paragraph>
      <Paragraph position="2"> The two types of referring intentions yield two distinct types of propositions. When the speaker does not intend any par.</Paragraph>
      <Paragraph position="3"> ticular aspect of the referent to be recognized by the hearer, the proposition expressed in this way is singular, that is, it does not contain any individual concept of the referent. Consequently, the referring expression chosen by the speaker (be it a proper name, a demonstrative, or even a definite description) is used as a rigid designator, which means that it picks out the same individual in all possible worlds where the referent exists. On the other hand, if a particular aspect of the referent is meant to be recognized by the hearer, then the individual concept corresponding to that aspect is part of the proposition expressed and should therefore be taken into account in evaluating the truth value of what is said. Thus, it is the speaker's referring intentions that determine whether or not he will use a definite description as a rigid designator (the modal distinction). Since referring intentions are represented in the planner, this is where the modal distinction belongs.</Paragraph>
      <Paragraph position="4"> Note that the two types of referring intentions can be described as intentions to place constraints on the way the hearer will be thinking of the referent. In Appelt and Kronfeld \[2\], this is generalized to other referring intentions -- for example, the intention that the hearer identify the referent.</Paragraph>
    </Section>
    <Section position="2" start_page="187" end_page="187" type="sub_section">
      <SectionTitle>
Referring Expressions
</SectionTitle>
      <Paragraph position="0"> Once the speaker decides what his referring intentions are, he must choose an appropriate referring expression. Usually, if a particular aspect of the referent is important, a suitable definite description is employed; otherwise a proper name or a demonstrative may be more useful. However, such a neat correlation between types of referring expressions and referring intentions may not happen in practice. In any case, as we shall see in the next section, the speaker's choice of a referring expression constitutes an implicit decision as to whether the denotation of the referring expression must coincide with the intended referent (the speech act distinction). The choice of referring expression is naturally made within the utterance generator, where the speech act distinction is represented.</Paragraph>
      <Paragraph position="1"> By way of summary, Table I shows how Donnellan's distinction, in its reinterpreted form, is related to a plan-based model of reference.</Paragraph>
    </Section>
  </Section>
  <Section position="8" start_page="187" end_page="188" type="metho">
    <SectionTitle>
RELEVANT DESCRIPTIONS
</SectionTitle>
    <Paragraph position="0"> Kripke and Searle \[12,15\] explain the referential use as a case in which speaker's reference is distinct from semantic reference. This leaves an important question unanswered: why must speaker's reference and semantic reference coincide in the attributive use? s Sometimes two definite descriptions are equally useful for identifying the intended referent, yet cannot be substituted for each other in a speech act. The description employed, besides being useful for identification, has to be relevant in some other respect. Consider the utterance: &amp;quot;New York needs more policemen.* Instead of &amp;quot;New York,&amp;quot; one might have used &amp;quot;The largest city in the U.S2 or &amp;quot;The Big Apple,&amp;quot; but &amp;quot;The city hosting the 1986 ACL conference needs more policemen&amp;quot; won't do, even though this description might be as useful in identifying New York as the others. The latter statement simply conveys an unwarranted implication.</Paragraph>
    <Paragraph position="1"> As a generalization, we may say that there are two senses in which a definite description might be regarded as relevant.</Paragraph>
    <Paragraph position="2"> First, it has to be relevant for the purpose of letting the hearer know what the speaker is talking about. 4 A description that is relevant in this sense may be called functionally relevant.</Paragraph>
    <Paragraph position="3"> S~eond, as the example above indicates, a description might exhibit a type of relevance that is not merely a referring tool.</Paragraph>
    <Paragraph position="4"> ~As redefined by the ~pcech act distinction.</Paragraph>
    <Paragraph position="5">  A description that is relevant in this noninstrumental sense might be called conversationally relevant.</Paragraph>
    <Paragraph position="6"> Every use of a definite description for the purpose of reference has to be functionally relevant. But not every such use has to be conversationally relevant. If indicating the referent is the only intended purpose, any other functionally relevant description will do just as well.</Paragraph>
    <Paragraph position="7"> In other cases, the description is supposed to do more than just point out the intended referent to the hearer. Consider the following examples:  In these examples, the speaker implicates (in Grice's sense} something that is not part of what he says. In (3), it is implicated that the man's happiness is due to his drinking. In (4), it is implicated that the main motivation for believing the murderer to be insane is that he committed such a brutal homicide. The implicature in (5) is that the only reason for giving the winner $10,000 is his victory in a particular race. In all these cases, what is implicated is some relationship between a specific characteristic of the referent mentioned in the description and whatever is said about that referent. In such cases, it does matter what description is chosen, since the relevance is both functional and conversational. No other description, even if it identifies equally well, can be as successful in conveying the intended implicature.</Paragraph>
    <Paragraph position="8"> The conversationally relevant description may not be mentioned explicitly, but rather inferred indirectly from the context. In the fingerprint example, the speaker uses the description, The man whose fingerprints these are, but the conversationally relevant description is nevertheless Smith's murderer. Thus, there are three general ways in which a speaker may  employ a referring definite description: 1. If the discourse requires no conversationally relevant description, any functionally relevant one will do. This covers all standard examples of the referential use, in which the sole function of the definite description is to indicate an object to the hearer.</Paragraph>
    <Paragraph position="9"> 2. If a conversationally relevant description is needed, the speaker may do either of the following: (a) Use the description explicitly. This is what is done in standard examples of the attributive use.</Paragraph>
    <Paragraph position="10"> (b) Use a different, functionally relevant description.</Paragraph>
    <Paragraph position="11">  The speaker can do so, however, only if the context indicates the aspect of the referent that corresponds to the conversationally relevant description. This explains the ambiguity of the fingerprint example. As the definite description uttered is only functionally relevant, its use appears to be referential. Yet, unlike the referential case, a conversationally relevant description is implied.</Paragraph>
    <Paragraph position="12"> In sum, when the description used is conversationally relevant, the speaker intends that the specific way he chose to do his referring should be taken into account in interpreting the speech act as a whole. Consequently, if the description fails, so does the entire speech act. On the other hand, if the description is only fimctionally relevant, the context may still supply enough information to identify the intended referent.</Paragraph>
  </Section>
class="xml-element"></Paper>