Prevention and Repair of Breakdowns in a Simple Task Domain

Brian Hansen, David G. Novick and Stephen Sutton

Center for Spoken Language Understanding
Oregon Graduate Institute of Science & Technology
P.O Box 91000, Portland, OR 97291-1000 USA
{brianh, novick, sutton}@cse.ogi.edu

Abstract

This paper identifies and explicates examples of prevention and repair of conversational breakdowns in a very simple domain consisting of sequences of letters. The evidence from the corpus suggests that conversants employ a number of specific repair-avoidance methods, and rely on acceptance and dis-acceptance to guide the interaction. These methods include managing turns to avoid confusion from overlapping speech, directing the other conversant to take a particular conversational action, and using a system of installments and feedback. Conversants may not always be able to determine unambiguously the cause of a perceived breakdown. Acknowledgments, for example, may be used to perform so many different functions that the meaning of a particular acknowledgment is typically ambiguous. Should a breakdown occur, conversants go through a sequence containing four possible stages of repair: trigger, diagnosis, prescription and treatment. The corpus contains examples of self-repair, prompted self-repair, other-repair, and delayed repair.

Introduction

Human-computer interaction, spoken or otherwise, can often be characterized as task-based: the human user of the computer is trying to accomplish some goal beyond, say, entertainment or social amusement. When looking for techniques to avoid miscommunication between computers and users, it seems appropriate to explore the behaviors that humans exhibit in robust task-based conversation. Unfortunately, many observable task-based conversations require extensive domain knowledge in order to find, much less to understand, instances of coping with conversational breakdown. This paper, then, identifies and explicates examples of prevention and repair of conversational breakdowns in a very simple domain. The abstractions we introduce for repair provide a basis for computational implementation of avoidance and repair of conversational breakdowns.

We address the problem of breakdowns in terms of underlying conversational models. That is, the brief sections of conversations that one normally identifies as "repairs" are simply the visible manifestation of longer-term, more subtle processes. Clark (1993) contended that the view that conversants monitor for problems in speech and then repair them is too narrow; repair work is only a part of a larger system and is used only as a last resort. In our studies we find this perspective useful for explaining the behavior of human conversants in carrying out a simple task. In particular, we see many instances of both problem prevention and repair in our letter sequence corpus (LSC) (Novick, Hansen & Lander, 1994). From this viewpoint, the problem of repair is really the problem of robustness: the use of mechanisms to cope gracefully with perceived or potential breakdowns in mutuality of belief. Conversants must maintain, as a conversation progresses, a degree of mutuality of knowledge (Clark & Marshall, 1981) sufficient for each participant; for knowledge to be mutual, the conversants must both know it and know that they both know it.

Why, if humans have these powerful underlying mechanisms for robustness, does conversation break down? Conversants appear to be satisfied with "sufficient" mutuality. Consequently, human conversants often verge on failing to achieve mutuality. An illustration of the tenuousness of mutuality is the fact that human conversants may not only unknowingly diverge in their interpretations of prior utterances, but may have different recollections as to what words were actually uttered. In telephone speech we ultimately share only the transient speech signal, and that only partially. It is not uncommon, for example, to hear exchanges in reference to some prior utterance, something like the following:

A: But you said x.
B: No, I said y.

Consequently, we claim that repair in conversation is not an all-or-nothing thing; conversants tolerate--or even rely on--large amounts of ambiguity in their exchanges. The extent of this ambiguity is a function of the conversants' tolerance of uncertainty of mutuality of belief. The interactions normally identified as "repairs" are simply a more explicit extension of a continuous process of building common ground in conversation.

In this paper, we describe the letter sequence domain, in which most ordinary domain-level semantics have been reduced to letters of the alphabet and their sequences. Using data from dialogues in this domain, we identify and characterize phenomena of prevention and repair of conversational breakdowns with brief excerpts from the corpus. The evidence from the corpus suggests that conversants employ a number of specific repair-avoidance tactics, rely on acceptance and dis-acceptance to guide the interaction, and, should a breakdown occur, go through a sequence containing four possible stages of repair: trigger, diagnosis, prescription and treatment.

Methodology

In an earlier study of human-human interaction in a simple task-oriented domain (Novick, Hansen & Lander, 1994), we gave subjects the assignment of collaborating face-to-face in what we call the letter sequence task, a task designed to mimic the structure of more complex collaborative tasks in which participants have incomplete knowledge. Each pair of subjects was given cards containing a sequence of 16 letters and blanks such that they could reconstruct the entire sequence by pooling their knowledge. They were asked to put the cards out of sight and to work together to reconstruct the complete sequence from memory. Figure 1 shows a sample of two subjects' letter sequences.

Subject L: O _ O S E _ A G F H C W _ E _ _ Z
Subject R: _ S O S _ X _ G F H _ _ L E B Y _
Figure 1: Sample letter sequences

There are several advantages in using this task to study human-human spoken interaction. First, the task is generally difficult enough that conversational breakdowns are not uncommon. The speech that it elicits is spontaneous and is not stilted in the same way that "read speech" can be. The task is essentially a collaboration between equals and is inherently a mixed-initiative situation. While the task centers on an extremely simple domain (what Henry Hamburger called the e. coli of dialogue), it mimics other more practical kinds of tasks where conversants have only partial information and must collaborate in order to complete the task successfully. Finally, the domain of the task lends itself to extending the experiment to study human-computer interaction, because the technology is already in place to recognize spoken letters at a level of accuracy above 90 percent.

The letter sequence task also has a number of disadvantages with respect to achieving understanding of human interactive behavior. In particular, the task is artificial and is performed under artificial conditions (subjects talk to a stranger either in the laboratory or over the telephone while the experimenter directs them). The task imposes fairly severe requirements on the memory capacity of the subjects such that completion of the task is not always possible. Furthermore, the task is inherently sequential; to the subjects it is more akin to collaborating to reconstruct a half-forgotten recipe than it is like collaboratively identifying its ingredients. Finally, although automatic spoken letter recognition is a well-established technology, many letters have common function words as homonyms ("I see" sounds like "I C", "Okay" sounds like "O K" and so on).

Four pairs of subjects were recorded performing the experiment twice. We used two 8mm camcorders to capture individual gestures and facial expressions and a 3/4" video recorder to capture the overall interaction. These recordings were the basis for a set of verbal and gestural transcripts, narratives, and specialized transcripts depicting the types of feedback the subjects provided and the relationship between direction of gaze and turn-taking. Transcripts may be found in (Novick, Hansen & Lander, 1994). In the examples discussed in the text, conversations from this corpus are identified with a designator like "F01.2:34-38," meaning face-to-face conversation, pair 01, trial 2, dialogue events 34 through 38.

We also collected and analyzed a set of telephone conversations in the same domain. Twelve pairs of subjects were recorded digitally; verbal transcripts were prepared from the audio files. The telephone corpus is not yet published. In the examples discussed in the text, conversations from this corpus are identified with a designator like "T52.1:35-41," meaning telephone conversation, pair 52, trial 1, dialogue events 35 through 41.

Humans are so good at dealing with ambiguity or minor breakdowns in mutuality that we often hardly notice them; and the variability of conversants' adaptations increases the difficulty of analyzing misunderstandings (Humphreys-Jones, 1987). Accordingly, we work from tapes and transcripts, using a mix of conversation-analytic and discourse-analytic techniques to identify patterns and produce corresponding models of robust interaction. We look at means of preventing breakdowns, at the role of positive and negative feedback in guiding conversations, and finally at the timing and contents of repair utterances.

Breakdown Prevention

In the LSC conversations we found several ways in which conversants work to avoid potential breakdowns and subsequent repair episodes. These methods include managing turns to avoid confusion from overlapping speech, directing the other conversant to take a particular conversational action, and using a system of installments and feedback.

Turn Management and Overlapping Speech

In face-to-face interaction the speech channel is a limited resource. In many circumstances, conversants make use of techniques to minimize and mitigate overlapping speech. In particular, speakers may signal their unwillingness to relinquish the floor by use of prosody, gestures, or turn-holding words. Hearers may indicate their desire for the turn-at-speech by close monitoring of transition relevance points (Sacks, Schegloff & Jefferson, 1974) and by use of filled pauses, or by outright interruption. Following an interruption attempt, the original speaker, or the new speaker, will often restart their utterance from the beginning, implicitly acknowledging that some of the meaning may have been lost during the overlapping speech (Goodwin, 1981).

Interestingly, we found several instances of overlapping speech that did not prompt correction or restarts. Some letter-task conversants were able to coordinate the rhythm of their interaction to the extent that they were able to "chant" together their sequences of letters or blanks, stopping when a mismatch or ambiguity occurred. While avoiding overlapping speech is often a good way of avoiding problems leading to breakdowns requiring repair, simultaneous speech may not necessarily require repair if it is closely coordinated and synchronized. The following example from the LSC illustrates this use of simultaneous speech. The conversants' speech overlaps but B indicates acceptance of A's presentation, presumably because he has understood that the overlapped utterances had convergent meanings.

Example 1: [F01.2:100-103]

A: Okay
A: //So that would have been "S"//
B: //So that would have been yours.//
B: Right

Directing Conversational Action

A second means of preventing repair is for one conversant to direct explicitly the other in how to respond. We find several instances in which subjects employed this strategy. In practice, experience in building working spoken language systems has led us to design dialogue models that direct the other's actions very explicitly (e.g., "Please say yes or no," and "Please wait for the tone") (Cole et al., 1993). Example 2 shows conversant B politely directing the other's action.

Example 2: [F1.1:2-12]

B: Okay. I guess you get to start, hunh?
A: Okay. uh, the first letter I had was V.

Installments and Feedback

The third and most pervasive technique for preventing breakdowns we found in the letter sequence corpus is the use of "installments" (Clark & Schaefer, 1989). Nothing in the instructions directs the subjects to use a close-coordination strategy in performing the task. It seems to be either an unconscious choice or an inherent calculation (Clark & Brennan, 1991) on the part of the subjects to break down the task into its smaller constituents and to act so as to achieve mutuality with regard to them. We would, for example, expect quite different behavior if the subjects were asked to perform the same task via e-mail.

By breaking the task into sub-sequence and letter installments, the conversants have the opportunity to give and to receive feedback about individual domain elements and to resolve conflicts or ambiguities as they occur. Clark and Schaefer's (1989) framework for evidence of understanding described actions that may take place following a successful perception of the other's contribution. Specifically, they identified five levels of acceptance of increasing "strength": continued attention, next relevant contribution, acknowledgment, demonstration, and display. Conversants often produce a sequence of acceptances in response to the other's antecedent contribution. In Example 3, conversant B accepts A's installment of a sub-sequence of letters by an acknowledgment followed by a next relevant contribution.

Example 3: [F1.1:21-26]

A: and then I had Y and S.
B: Okay. and then I had P.

With these brief examples, we have tried to show a range of techniques employed by conversants to prevent breakdowns in conversation. In each of the examples the conversants, sometimes overcoming possible confusion, reach a coherent conclusion in their exchanges. In particular, the use of conversational installments establishes expectations upon which the participants can rely for disambiguation. A central feature of the use of installments for feedback is the reliance on acceptance--and dis-acceptance--to guide the interaction. We now turn to the points in the conversation where participants choose either to accept a presentation or to initiate a repair.

Diagnosis: Between Acceptance and Repair

With each presentation--in fact, within each presentation--comes the opportunity to accept or dis-accept. That is, throughout and at the end of each utterance a hearer can indicate (a) continuing acceptance, (b) dis-acceptance, perhaps by interrupting, or (c) degrees of acceptance and dis-acceptance. This dynamic process is central to the tailoring of verbal production to the particular requirements of the hearer, such as shown by Schober and Clark (1989) and Oviatt and Cohen (1988). In this border territory between acceptance and repair are found the phases of conversational interaction that are the mechanisms common to both.

In looking at acceptances, it would be useful to consider their control functions independently of the domain acts that they accomplish. For these purposes, a study of relevant contributions as acceptances is problematic because they serve other functions beyond "mere" acceptance; they add new domain knowledge to the dialogue and require domain inference to interpret. The LSC, like most other corpora, contains few instances of display or demonstration acceptance levels. Accordingly, in this study, we focus on acknowledgments because they are independent of domain knowledge, relatively easy to identify, and sufficiently frequent for useful analysis.

Acknowledgments, and acceptances in general, are ubiquitous in the LSC. In some corpora (e.g., Marshall & Novick, 1995) it is not unusual to find that a surprisingly large percentage (10-25 percent) of utterances begin with or consist entirely of acknowledgments. Why so many? After all, one might reason that these explicit acknowledgments take extra time to say as well as the time it takes to coordinate their production and perception. Would dialogue not be more efficient if each conversant simply said the relevant things and dispensed with acknowledgments? One reason is that between a thing being said and its being understood as being mutual, a lot can go wrong. Another reason is that acknowledgments perform many different functions in dialogue. In an earlier study, we identified eleven functions that acknowledgments may perform (Novick & Sutton, 1994a, 1994b). These functions include:

Since acknowledgments may be used to perform so many different functions, the meaning of a particular acknowledgment is typically ambiguous. Additionally, when a conversant accepts a presentation, she does not necessarily provide unconditional acceptance. She might want to initiate some kind of repair anyway indicating (perhaps via prosody) that there remains some lingering uncertainty, doubt or disagreement with what was being nominally accepted.

The situation retains a great deal of ambiguity even if conversant A correctly perceives the other's response. To some extent, the mutuality of what A said depends upon what happens in the rest of the dialogue. Mutuality is always provisional and amenable to modification at a later time. Practically, however, if all has gone well through two exchanges, then it is often reasonable to assume the mutuality of the original utterance. In other words, A's belief that it is mutual depends upon what A perceives to be B's reaction to A's response to what B originally said.

Conversely, it may not always be possible to determine unambiguously the cause of a perceived breakdown. Dis-acceptance may signal problems at any level. If A perceives that B didn't accept A's contribution, then, for example, did B:

Although not described in the context of acceptance, the central aspects of these conditions were formally modeled by Reilly (1987). Each condition represents a distinct process for miscommunication. When the result of one of these processes is a misunderstanding that outstrips a conversant's ability for prevention or tolerance of mistakes, then the interaction turns to explicit repair.

Places and Stages of Repair

Breakdowns in mutuality and moves made to resolve those breakdowns may occur at virtually any point within a conversation. For instance, the most recent contribution by the other may reveal that the other conversant holds conflicting beliefs with respect to some topic that was seemingly resolved many turns ago. Or, the implications of the other conversant's prior actions may suddenly be recognized and acted upon. Similarly, the breakdown may have been recognized earlier but not viewed as important enough to resolve, or may have been deferred in the interest of efficiency.

The transience of spoken behavior and the difficulties and conversational "expense" (Clark & Brennan, 1991) of effecting resolution of problems whose source is too far in the past suggest that the sooner that problems are resolved the better. Schegloff, Jefferson & Sacks (1977), and van Lier (1988) classify repairs in terms of the turn in which they appear and how the work of initiating repair is allocated among the participants. Here we take a more functional approach. We begin by examining repair episodes that take place immediately (self-repair), and those that take place later, whether due to delayed problem recognition or by a choice on the part of the conversants to defer repair. We follow this by briefly identifying the components that make up a typical repair episode in the LSC.

Self-Repair

We found several instances of self-repair in the letter sequence corpus. In addition to straightforward self-repair episodes, we found instances of what might otherwise be considered self-repair that could be seen to be either cued or reinforced by non-verbal feedback (wincing or looking away) by the other. Similarly, a repair episode was sometimes triggered by overly weak acceptance. Additionally, we found several instances of delayed repair, in which problems with memory or in comprehending the implications of a breakdown caused a delay in repair initiation, often leading to extended repair subdialogues.

The LSC conversations in the face-to-face condition contained few cases of "pure" (i.e., un-cued) self-repair; these occurred at a higher rate in the telephone condition. In Example 4, conversant B, formulating a response to A's question, performs what must be a pure self-repair because the part of the utterance under repair is at that point yet unuttered. Similarly, in Example 5, conversant B restarts his utterance when he apparently realizes that his dialogue plan is unformed.

Example 4: [F1.2:150-162]

A: Was that a blank?
B: ((to self)) I had A F G no I, I had A F G H C W all in a row.

Example 5: [T47.1:6-8]

B: Start with ... ((pause)) I'm a little confused here as to where we need to start.

Much more prevalent in the face-to-face condition were cases of repairs cued by non-verbal actions of the "hearing" conversant. Example 6 (presented in a detailed transcript in Figure 2) is a case where conversant A initiates an apparent self-repair:

Example 6: [F04.2:9-21]

A: H C F no, A! C F ...

Review of the non-verbal actions set out in the transcript suggests that A's repair was actually triggered by B's action of looking away from A; one beat after B shifts his gaze, A begins the repair. Presumably B was communicating a higher level of uncertainty about A's utterance.

Conversant A Conversant B
FACE/BODY HANDS EYES VERBAL VERBAL EYES HANDS FACE/BODY
  
  
  
  
  
to
  
  
  
counting away H
  
  
  
  
  
to
C
  
  
  
  
  
  
  
  
away
  
  
  
F
  
  
  
  
  
  
thumb back away No
  
  
  
  
  
counting with thumb
  
A (!)
  
  
  
  
  
to C
  
  
  
  
  
  
F
  
  
  
  
  
counting with hands
  
H
  
  
  
  
  
  
C
  
  
  
  
  
  
W
  
  
  
  
  
  
  
  
  
close
  
  
Figure 2: Transcript for Example 6 [F04.2:9-21]

Immediate and Delayed Repairs

Frequently conversants will conduct an immediate repair of an apparent error. In Example 7, conversant B, after a couple of uncued self-repairs, receives a dis-acceptance and presentation from A that leads B to clarify his utterance. In Example 8, Conversant A triggers an immediate repair by dis-accepting B's presentation.

Example 7: [T:60.0:1-7]

B: What- What's your first two blank- uh your first two letters?
A: I don't have numbers. I have letters.
B: Well, I mean first two letters.

Example 8: [T47.2:54-60]

A: Uh huh ((rising tone)).
B: And the last one is N.
A: N?
B: N. Yes.
A: Okay.

The most common form of delayed repair found in the LSC involved deferring repair until the end of the current sub-sequence or until the end of the entire sequence had been reached. Example 9 shows the case of conversant A deferring the initiation of repair until B had finished her entire sub-sequence. Example 10 depicts the more common situation in which the conversants defer repair until the end of the full sequence; conversant B could have initiated a repair when he first detected a lack of mutuality but instead deferred the repair to a major domain segment boundary.

Example 9:[F1.2:111-127]

B: And then there was, uh, A G F H C W and then I had a blank.
A: O F G? or O G F?

Example 10:[F4.2:28-38] (at the end of the sequence)

B: There's something that didn't sound right in there to me. Let's try that again.

These examples show that repairs can occur in a variety of places within the interaction. Pure self repairs can occur anywhere because they result from an internal process of the speaker. For other repairs, the point of occurrence may be a function of the speaker's perception of and willingness to accommodate signals from the hearer that the current presentation is problematic. Where the speaker is immediately responsive, she will produce a cued self-repair. Otherwise, the hearer may produce an interruption or a delayed repair initiation.

Stages of Repair

In our analysis of repair episodes in the letter sequence corpus we identified a series of stages that conversants go through in resolving breakdowns. While not all repair episodes exhibit all of the stages we identify, virtually all contain some of the following key elements:

Example 11 presents a repair episode that displays all four stages. Conversant A triggers a repair by dis-accepting B's contribution. After a pause, B begins diagnosing the cause of A's problem. After B rules out her first try at diagnosis, she begins a second attempt, during which she has an insight into what might be the cause of the problem and restarts. A agrees with B's reformulated diagnosis and offers as a prescription that B offer the next letter. B commences (during a minor speech overlap) the treatment of the problem, but starts from an earlier position in the sequence than A suggested. Note that the subjects in this conversation did not successfully resolve this problem, but shifted from the belief that A had forgotten two letters to the belief that both conversants sequences were blank at those positions in the sequence.

Example 11: [F1.1:125-142]

B: P
A: mmm...
((pause))
B: The P was a blank for you though.
A: M-hunh.
B: So the next two... Oh, those are the two that you're having a hard time (with)?
A: Yeah. So and then what was the next (one) //you had? //
B: //A and// F after the P. Then there are two blanks.

The trigger stage, in which a conversant signals the possible initiation of a repair, can be carried out through explicit dis-acceptance, request for repair, or weaker-than-expected acceptance.

The corpus contains examples of each of these actions. In Example 11, above, the repair was triggered by an explicit dis-acceptance. In Example 12, conversant A produces a request for repair that triggers a diagnosis.

Example 12: [T57.0:55-59]

B: N, then I have G, S.
((pause))
A: F like Frank?
B: S as in Sam.
A: Sam ...

In telephone conversation, continued attention is indistinguishable from non-attention. Consequently, silence is ambiguous because it neither accepts nor dis-accepts. In Example 13, silence serves as an instance of weaker-than-expected acceptance, triggering a repair. In this case, B does not take up A's request, and A goes on to self-repair.

Example 13: [T49.0:50-58]

A: Oh! I've got (gee whiz) ... uh, ohh boy. N A. Yeah, N A.

If repair is deferred, or is triggered after a delay, conversants may have to provide context for the diagnosis:

Example 14: [T60.0:119-123]

B: And then //where's ...//
A: //What's// happened to your blank C blank at the beginning of your //sequence?//
B: //Well// I filled in your numbers with it.
A: That third one there... that didn't sound right.

Similarly the prescription may specify a context:

Example 15: [F03.2:60-65]

A: Okay um do you want to start at the beginning again or ...
B: ((laughs)) Okay.

We also posit a stage "0" of repair that involves detection. That is, prior to diagnosis the conversant detects a discrepancy between what is heard and what is expected. The evidence for this stage is weak because it is a mental rather than a behavioral construct. Such a stage could, however, represent an integration of the prevention and repair aspects of dialogue. In particular, a detection stage would trigger prevention behaviors such as the "cueing" that leads to self-repair. The rather subtle gestural behaviors, such as facial expression, that we suggest are repair cues may be routine signals of the detection process at work.

Discussion

The letter-sequence task is simple, collaborative, mixed-initiative, and has a strong emphasis on mutuality of knowledge. It was selected to keep the required domain knowledge simple enough so that it was not a confusing factor, and to eliminate the possibility that the subjects had prior knowledge of the domain that was inaccessible to the researchers.

More realistic domains for systems supporting language-based human-computer interaction are likely to share some of these characteristics. In particular, applications will probably be task-based, collaborative, and have relatively simple vocabularies and grammars. While currently deployed technology for spoken language systems typically does not use mixed initiative, prototype systems with mixed initiative are now being developed (see, eg., Fanty et al., 1995). Indeed, lack of support for mixed initiative is usually regarded as a shortcoming.

It is possible that the nature of the letter sequence task itself influenced the strategies used by the conversants. Moreover, conversants may employ domain-specific repair actions in other domains. Consequently, generalization of these results depends on their application to new domains.

The discourse-analytic abstractions developed in this study extend the representations of conversational control acts that have been successfully modeled in agent-based simulations (eg., Novick & Ward, 1993). Further work in this area should include computational representation of the breakdown prevention and repair behaviors that we have identified.

In this paper, we have described the letter sequence task, an extremely simplified domain for studying the mechanics of human-human spoken interaction. Using data from dialogues in this domain, we have identified and characterized phenomena of prevention and repair of conversational breakdowns. The evidence from the corpus suggests that conversants employ a number of specific repair-avoidance tactics, rely on acceptance and dis-acceptance to guide the interaction, and, should a breakdown occur, go through a sequence containing four possible stages of repair: trigger, diagnosis, prescription and treatment.

Acknowledgments

This work was funded by National Science Foundation awards IRI-9314959 and GER-9354959, the Advanced Research Projects Agency, and the Center for Spoken Language Understanding. Additionally, we thank Karen Ward and the reviewers for their comments.

References

Clark, H. 1993. Managing problems in speaking, Proceedings ISSD-93, International Symposium on Spoken Dialogue: New Directions in Human and Man-Machine Communication, Nov. 10-12, 181-184.

Clark, H. and Brennan, S. 1991. Grounding in communication, Shared Cognition: Thinking as Social Practice, APA Books.

Clark, H. and Marshall, C. 1981. Definite Reference and Mutual Knowledge, Elements of Discourse Understanding. Cambridge: Cambridge University Press, 10-63.

Clark, H. and Schaefer, E.F. 1989. Contributing to discourse, Cognitive Science, 13, 259-294.

Cole, R., Novick, D., Fanty, M., Sutton, S., Hansen, B. and Burnett, D. 1993. Rapid prototyping of spoken-language systems: The Year 2000 Census project. Invited paper, Proceedings of the Conference on Spoken Language Systems, Tokyo, Japan, November, 1993, 19-23.

Fanty, M., Sutton, S., Novick, D., and Cole, R. 1995. Automated appointment scheduling, ESCA Workshop on Spoken Dialogue Systems, Vigso, Denmark, May, 1995, 144-47.

Goodwin, C. 1981. Conversational organization. New York: Academic Press.

Humphreys-Jones, C. 1987. The structure of misunderstandings. In Reilly, R. (ed.), Communication failure in dialogue and discourse. Amsterdam: Elsevier Science Publishers.

Marshall, C., and Novick, D. 1995. Conversational effectiveness in multimedia communications, Information Technology & People, 8(1), 54-79.

Novick, D., Hansen, B. and Lander, T. 1994. Letter sequence dialogues, Technical Report No. CS/E 94-007, Oregon Graduate Institute of Science & Technology.

Novick, D. and Sutton, S. 1994a. An empirical model of acknowledgment for spoken-language systems, Proceedings of the 32nd Annual Meeting of the Association for Computational Linguistics (ACL'94), 96-101.

Novick, D. G. and Sutton, S. 1994b. Acknowledgment acts in task-based dialogue. Technical Report No. CSE94-01, Department of Computer Science and Engineering, Oregon Graduate Institute of Science & Technology.

Novick, D., and Ward, K. 1993. Mutual beliefs of multiple conversants: A computational model of collaboration in air traffic control. Proceedings of AAAI'93, Washington, DC, July, 1993, 196-201.

Oviatt, S., and Cohen, P. 1988. Discourse structure and performance efficiency in interactive and non-interactive spoken modalities. Technical Note 454, SRI International.

Reilly, R. 1987. Types of communication failure in dialogue. In Reilly, R. (ed.), Communication failure in dialogue and discourse. Amsterdam: Elsevier Science Publishers.

Sacks, H., Schegloff, E.A. and Jefferson, G. 1974. A simplest systematics for the organization of turn-taking for conversation, Language, 50, 696-735.

Schegloff, E.A., Jefferson, G., Sacks, H. 1977. The preference for self-correction in the organization of repair in conversation, Language, 53(2), 361-382.

Schober, M., and Clark, H. 1989. Understanding by addressees and hearers, Cognitive Psychology, 21, 211-232.

van Lier, L. 1988. The classroom and the language learner. Ch. 7.: The organization of repair in second language classrooms. Essex, Engl.: Longman, 180-212.


This paper appeared as Hansen, B., Novick, D., & Sutton, S. (1996). Prevention and repair of breakdowns in a simple task domain. Working notes, AAAI 96 Workshop on Detecting, Repairing, and Preventing Human-Machine Miscommunication, August 4, 1996, Portland, OR, 5-12.

__________
DGN, October 10, 1997