<?xml version="1.0" standalone="yes"?>
<Paper uid="J86-4002">
  <Title>REFERENCE IDENTIFICATION AND REFERENCE IDENTIFICATION FAILURES</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 INTRODUCTION
</SectionTitle>
    <Paragraph position="0"> Cohen, Perrault, and Allen (1981) argued that &amp;quot;... users of question-answering systems expect them to do more than just answer isolated questions - they expect systems to engage in conversation. In doing so, the system is expected to allow users to be less than meticulously literal in conveying their intentions, and it is expected to make linguistic and pragmatic use of the previous discourse.&amp;quot; Following in their footsteps, we want to build robust natural language processing systems that can detect and recover from miscommunication. The development of such systems requires a study on how people communicate and how they recover from miscommunication. The study of miscommunication is a necessary task for building natural language understanding systems since any computer capable of communicating with humans in natural language must be tolerant of the complex, imprecise, or ill-devised utterances that people often use. This paper summarizes the results of a dissertation (Goodman 1984) that investigated the kinds of miscommunication that occur in human communication, with a special emphasis on reference problems: problems a listener has in determining about whom or what a speaker is talking. To cope with such problems, we proposed an algorithm for extending the reference paradigm. We have also implemented computer programs that demonstrate how one could solve these problems in a natural language understanding system.</Paragraph>
    <Paragraph position="1"> Our current research (Sidner et al. 1981, 1983) assumes most dialogue as being cooperative and goal directed: we assume that a speaker and listener are working together to achieve a common goal. In order for the listener to interpret utterances, he must identify the underlying plan or goal that the utterances reflect (Cohen 1978, Allen 1979, Sidner and Israel 1981, Sidner 1985, Carberry 1985, Litman 1985, Pollack 1986). This plan, however, is rarely obvious at the surface sentence level. A central process, therefore, in the interpretation of utterances is the transformation of sequences of complex, imprecise, or ill-devised utterances into well-specified plans that might be carried out by dialogue participants. Within this process, miscommunication can occur. In this paper, we are particularly concerned with cases of miscommunication from the hearer's viewpoint, such as when the hearer is inattentive to, confused about, or misled about the intentions of the speaker.</Paragraph>
    <Paragraph position="2"> In ordinary exchanges speakers usually make assumptions regarding what their listeners know about a topic of discussion. They will leave out details thought to be superfluous (Appelt 1981C/ McKeown 1983). Since the speaker really does not know exactly what the listener knows about a topic, it is easy to make statements that can be misinterpreted or not understood by the listener Copyright1986 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/86/040273-305503.00 Computational Linguistics, Volume 12, Number 4, October-December 1986 273 Bradley A. Goodman Reference Identification and Reference Identification Failures because not enough details were presented. One principal source of trouble is the descriptions constructed by the speaker to refer to actual objects in the world. A description can be, for a given listener, either imprecise, confused, ambiguous, or overly specific. In addition, it might be interpreted under the wrong context (which can cause one of the problems with the description to occur or can cause the description to successfully refer when it should not have). As a result, reference identification errors 2 occur; the listener cannot determine what object is being described. The descriptions that cause reference identification failure are a type of &amp;quot;ill-formed&amp;quot; input. The blame for ill-formedness may lie partly with the speaker and partly with the listener. The speaker may have been sloppy or may not have taken the hearer into consideration. The listener may be remiss, unwilling to admit he can't understand the speaker, or unwilling to ask the speaker for clarification. It may even be the case that the listener does not know that he has misunderstood the speaker.</Paragraph>
    <Paragraph position="3"> The interactions that can occur among the speaker's description, the context of the communication, and the listener's view of the world, as well as the listener's own abilities, especially in a task-oriented environment, all contribute to make the reference task more complicated.</Paragraph>
    <Paragraph position="4"> Our work provides a new way to look at reference that involves a more active, introspective approach to repairing communication. It redefines the notion of finding a referent since previous paradigms have proven inappropriate in the real world (see Section 4 for a detailed discussion).</Paragraph>
    <Paragraph position="5"> We introduce a new process to reference called negotiation that is used during the reference task to take into account all the language and perceptual knowledge people have about the world, especially when reference fails. We illustrate this process by introducing a new computational model for the reference process called FWIM, for &amp;quot;Find What I Mean&amp;quot;. In addition, we develop a theory of the use of extensional descriptions that will help explain how people successfully use imperfect descriptions. This theory is called the theory of extensional reference miscommunication.</Paragraph>
    <Paragraph position="6"> The last part of this section provides an introduction to the domain of our work and outlines the methodology used. We also present a description of other relevant research in this domain. Section 2 of this paper briefly highlights some aspects of normal communication and then provides a general discussion on the types of miscommunication that occur in conversation, concentrating primarily on reference problems and motivating many of them with illustrative protocols. The protocols demonstrate the complexity of the reference process and help illuminate the kinds of knowledge sources people consult when performing reference. Section 3 describes those knowledge sources in more detail, providing information about the language and physical knowledge that people use to perform reference identification and to recover from reference failure. Section 4 presents initial solutions to some of the problems of miscommunication in reference. Motivated there is a partial implementation of a reference mechanism that attempts to overcome many reference problems. Finally, we conclude in Section 5 with a summary and suggestions for future research.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
1.1 THE DOMAIN AND METHODOLOGY
</SectionTitle>
      <Paragraph position="0"> We are following the task-oriented paradigm of Grosz (1977) since * it is easy to study (through videotapes), * it places the world in front of you (a primarily extensional world), and * it limits the discussion while still providing a rich environment for complex descriptions.</Paragraph>
      <Paragraph position="1"> The task chosen as the target for the system is the assembly of a toy water pump. The water pump is reasonably complex, containing four subassemblies built from plastic tubes, nozzles, valves, plungers, and caps that can be screwed or pushed together. A large corpus of dialogues concerning this task was collected by Cohen (see Cohen 1981, 1984; Cohen, Fertig, and Starr 1982). These dialogues contained instructions from an &amp;quot;expert&amp;quot; to an &amp;quot;apprentice&amp;quot; that explain the assembly of the toy water pump. Both participants were working to achieve a common goal - the successful assembly of the pump.</Paragraph>
      <Paragraph position="2"> This domain is rich in perceptual information, allowing for complex descriptions of its elements. The data provide examples of imprecision, confusion, and ambiguity as well as attempts to correct these problems.  The following exchange exemplifies one such situation. In it, E is instructing A to assemble part of the water pump. E and A are communicating verbally but neither can see the other. (The bracketed text in the excerpt tells what was actually occurring while each utterance was spoken.) Notice the complexity of the speaker's descriptions and the resultant processing required by the listener. This dialogue illustrates that listeners * repair the speaker's description in order to find a referent, null * repair their initial reference choice once they are given more information, and * can fail to choose a proper referent.</Paragraph>
      <Paragraph position="3"> In Line 7, E describes the two holes on the BASEVALVE as &amp;quot;the little hole&amp;quot;. A must repair the description, realizing that E doesn't really mean &amp;quot;one&amp;quot; hole but is referring to the &amp;quot;two&amp;quot; holes. A apparently does this since he doesn't complain about E's description and correctly attaches the BASEVALVE to the TUBEBASE.</Paragraph>
      <Paragraph position="4"> Figure lb. Configuration of the pump after the TUBEBASE is attached to the MAINTUBE (Line 10).</Paragraph>
      <Paragraph position="5"> In Line 13, A interprets &amp;quot;a red plastic piece&amp;quot; to refer to the NOZZLE. When E adds the relative clause &amp;quot;that has four gizmos on it,&amp;quot; A is forced to drop the NOZZLE as the referent and to select the SLIDEVALVE. In Lines 17 and 18, E's description &amp;quot;the other - the open part of the main tube, the lower valve&amp;quot; is ambiguous, and A selects the wrong site, namely the TUBEBASE, in which to insert the SLIDEVALVE. Since the SLIDEVALVE fits, A doesn't detect any trouble. Lines 20 and 21 keep A from thinking that something is wrong because the part fits loosely. In Lines 27 and 28, A indicates that E has not given him enough information to perform the requested action. In Line 30, A further compounds the error in Line 18 by putting the SPOUT on the TUBEBASE.</Paragraph>
      <Paragraph position="6"> Excerpt 1 (Telephone) E: 1. Now there's a blue cap \[A grabs the TUBEBASE\]  2. that has two little teeth sticking 3. out of the bottom of it.</Paragraph>
      <Paragraph position="7"> A: 4. Yeah.</Paragraph>
      <Paragraph position="8"> E: 5. Okay. On that take the 6. bright shocking pink piece of plastic \[A takes BASEVALVE\] 7. and stick the little hole over the teeth.</Paragraph>
      <Paragraph position="9">  \[A starts to install the BASEVALVE, backs off, looks at it again and then goes ahead and installs it\] A: 8. Okay.</Paragraph>
      <Paragraph position="10"> E: 9. Now screw that blue cap onto 10. the bottom of the main tube.</Paragraph>
      <Paragraph position="11"> \[A screws TUBEBASE onto MAINTUBE\] A: 11. Okay.</Paragraph>
      <Paragraph position="12"> E: 12. Now, there's a13. a red plastic piece \[A starts for NOZZLE\] 14. that has four gizmos on it.</Paragraph>
      <Paragraph position="13"> \[A switches to SLIDEVALVE\] A: 15. Yes.</Paragraph>
      <Paragraph position="14"> E: 16. Okay. Put the ungizmoed end in the uh 17. the other-the open 18. part of the main tube, the lower valve.</Paragraph>
      <Paragraph position="15"> \[A puts SLIDEVALVE into hole in TUBEBASE, but E meant OUTLET2 of MAINTUBE\] A: 19. All right.</Paragraph>
      <Paragraph position="16"> E: 20. It just fits loosely. It doesn't 21. have to fit right. Okay, then take 22. the clear plastic elbow joint.</Paragraph>
      <Paragraph position="17"> \[A takes SPOUT\] A: 23. All right.</Paragraph>
      <Paragraph position="18"> E: 24. And put it over the bottom opening, too. \[A tries installing SPOUT on TUBEBASE\] A: 25. Okay.</Paragraph>
      <Paragraph position="19"> E: 26. Okay. Now, take the-A: 27. Which end am I supposed to put it over? 28. Do you know? E: 29. Put the-put the-the big end30. the big end over it.</Paragraph>
      <Paragraph position="20"> \[A pushes big end of SPOUT on TUBEBASE, twisting it to force it on\] The example illustrates the complexity of reference indentification in a task-oriented domain. It shows that people do not always give up when a speaker's description isn't perfect (or isn't readily assimilable for them), but that they try to plow ahead anyway. The rest of this paper formalizes the kinds of problems that occur during reference and then extends the reference paradigm to get around many of the problems. Computational Linguistics, Volume 12, Number 4, October-December 1986 275 Bradley A. Goodman Reference Identification and Reference Identification Failures</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>