<?xml version="1.0" standalone="yes"?>
<Paper uid="J86-4002">
  <Title>REFERENCE IDENTIFICATION AND REFERENCE IDENTIFICATION FAILURES</Title>
  <Section position="3" start_page="0" end_page="0" type="relat">
    <SectionTitle>
1.2 RELATED WORK IN REFERENCE AND MISCOMMUNI-
CATION
</SectionTitle>
    <Paragraph position="0"> There are two major pieces of work in AI literature that laid the foundation for our research: those in reference and those in miscommunication.</Paragraph>
    <Paragraph position="1"> Cohen (1981, 1984) presents a detailed analysis of the pragmatics of reference and the effects of different modalities of communication. His work was a major starting point of this research. It showed that it was reasonable to consider reference identification as separate from the whole process of language understanding instead of being too intimately tangled to consider on its own. There is evidence presented by Cohen (1981, 1984) that a speaker attempts as a separate step in his overall plan of communication to get a hearer to identify a referent. He provided grounds for an IDENTIFY action by illustrating particular requests to identify from his water pump protocols. For example, utterances like &amp;quot;'Notice the two side outlets on the tube end&amp;quot; or &amp;quot;'Find the rubber ring shaped like an O&amp;quot; showed that the speaker wanted the hearer to perform some kind of action.</Paragraph>
    <Paragraph position="2"> That action is the IDENTIFY act, which is to search the world for a referent for the speaker's description (and thus identify it). Cohen also showed that the hearer's response to a request to identify provided further evidence. He pointed out excerpts in the protocols where hearers responded to a request to identify with a confirmation that the identification had actually occurred (e.g., &amp;quot;Got it.&amp;quot;). Cohen went on to show how reference fits into a plan-based theory of communication.</Paragraph>
    <Paragraph position="3"> The reference paradigm we followed was closest to that developed by Grosz (1977). Her basic reference identification paradigm was similar to that of many others in the past (e.g., Winograd 1971, Woods 1972): put the speaker's description into a searchable form (i.e., parse and semantically interpret the speaker's description) and then use that form as a pattern that can be compared against objects (i.e., the possible referents) in the world. A referent is found when a match occurs between the pattern and one or more of the objects. The pattern and a target referent match each other if all the attributes specified in the pattern exactly fit the corresponding attributes in the target. There is variability in each of the past reference schemes in what pattern is generated, how the world is represented, and how the actual search progresses, but the general scheme remain.s the same. Success in all cases occurs if and only if a perfect match exists between all the pattern's attributes and the corresponding attributes on a target. Grosz's reference mechanism departed from past works by introducing the notion of focus. Focus provides a better way to resolve referents by constraining the search space. For definite noun phrases, the choice of possible referent candidates is guided by the focus mechanism. The information provided in the definite noun phrase itself (i.e., by the head noun and modifiers) is used to distinguish the referent from other objects in focus. Grosz showed how both the surrounding non-linguistic environment and the global context of preceding discourse are part of focus and how it is used to resolve definite noun phrases.</Paragraph>
    <Paragraph position="4"> Grosz (1977:161) also proposed the need for inexact matching in the reference process should something go wrong: The retrieval component can fail to find such a match even though for most people the noun phrase suffices to identify an object .... Alternatively, more than one object may match, but the ambiguity may not matter for the purposes of the utterance. The problem in either case is to determine the nature of the mismatch and whether it matters .... The focus mechanism provides one crucial element for deciding about inexact matches. It separates those items that are in the focus of attention from all other known items. If an exact match cannot be found in focus, it is reasonable to ask if any of the items in focus come close to matching the description of the noun phrase (the question of what is close is the other crucial element in such decisions) and if so which is closest.</Paragraph>
    <Paragraph position="5"> Ringle and Bruce (1981) present a survey of numerous types of miscommunication in conversation. They point out problems across a wide spectrum of dialogue types and situations. Two primary ways that conversation fail are described by them. The first one, input failure, occurs when the listener is unable to form a complete or at least coherent interpretation for an utterance. Input failure can occur due to such causes as misinterpretation of a single word, incorrect resolution of a referential term, and misplacement of a negation. Such failures cause the listener to misunderstand without weakening the listener's comprehension of the overall context of the communication (making the failures local in nature). The second way that Ringle and Bruce say that people fail, model failure, happens when the listener cannot incorporate the inputs into a coherent belief model as intended by the speaker. The problem could be due to an input failure when information is lost that is needed to assimilate the speaker's utterances into the belief model. It can also occur when a listener does not have sufficient background knowledge, has a different thematic emphasis than the speaker, or fails to make the proper inference (or any at all) from the speaker's input. Ringle and Bruce describe repair techniques that often occur between the listener and speaker when a failure occurs. Such repairs are usuany initiated by the listener providing a failure cue (e.g., recapitulating the speaker's important points) to the speaker to indicate possible trouble. The repairs primarily require action by both the listener and the speaker.</Paragraph>
    <Paragraph position="6"> Sometimes the dialogue situation affects the ability of the listener to provide such cues. For example, in a teacherstudent relationship, it is hard for the student to interrupt the teacher's lecture/conversation to initiate a repair due to a mistake the student feels has occurred. In other conversational settings, such interruptions are easier.</Paragraph>
    <Paragraph position="7"> McCoy (1985a, 1985b) focuses on a particular class of communication problems. She considers misconceptions about the objects modelled by a system in its 276 Computational Linguistics, Volume 12, Number 4, October-December 1986 Bradley A. Goodman Reference Identification and Reference Identification Failures knowledge base. She is concerned with discrepancies between the beliefs of the system and that of the user as seen in a system/user dialogue. Her work concentrates on two kinds of misconceptions about the properties of an object: misclassification and misattribution.</Paragraph>
    <Paragraph position="8"> Misclassification occurs when one classifies an object incorrectly. For example, a person may think that whales are fish when in fact they are mammals (McCoy 1985b:17). McCoy called the way to correct this problem the like-super strategy since an expert may believe that the user misclassified the misconception object (whale) because it is similar to the posited superordinate (fish). She defines two other kinds of misclassifications that her system can detect: Like-Some-Super and No-Support. Like-Some-Super occurs when the expert believes a user wrongly classified an object because it is like some subclass of the posited superordinate. For example, a whale may be viewed by someone as a fish because they think that a whale is like a shark, and a shark is a fish (McCoy 1985b:24). No-Support occurs when the system can find no support in the user model for the misclassification. McCoy' s system simply denies the incorrect information in that case and provides the correct information.</Paragraph>
    <Paragraph position="9"> Misattribution is the second class of misconceptions with which McCoy deals. They occur when the user wrongly attributes a property to an object that the object doesn't have. One reason that misattribution can occur is that the user either has confused the object with one the user thinks is a similar object or has made a bad analogy from a similar object (the &amp;quot;Wrong Object&amp;quot; strategy). McCoy presents an example where the user attributes the &amp;quot;high liquidity&amp;quot; property of a money market fund to a money market certificate. Another reason that misattribution can occur is that the user attributes to an object a related property instead of the actual one (the &amp;quot;Wrong Attribute&amp;quot; strategy). An example that McCoy presents for this strategy occurs when the user talked about the &amp;quot;interest&amp;quot; on the stock but really meant the &amp;quot;dividend.&amp;quot; The correction in that case is the substitution of the proper property for the incorrect one. The last case of misattribution that McCoy considers is No-Support. It occurs when the expert can find no support for the misattribution in his model of the user. In that case, McCoy's system denies the incorrect information and asserts the correct information.</Paragraph>
    <Paragraph position="10"> McCoy's work demonstrates the power of representing objects using a taxonomic knowledge base that indicates an object's superordinates and subtypes, and its attributes and their values. That paradigm allows her to notice several classes of user's misconceptions and to correct them. Her solutions blend in nicely with the relaxation mechanism motivated and described in this paper.</Paragraph>
  </Section>
class="xml-element"></Paper>