MITRE Corporation
luperfoy@mitre.org
CHI '97 Workshop: Speech User Interface
Design Challenges
By "repair dialogue" we mean a meta-dialogue in which the primary dialogue on the topic of the domain of discourse is suspended while a discussion on the topic of the dialogue itself, specifically some trouble in the dialogue, takes place. For example, for an information retrieval dialogue system to which the user has issued the spoken command, "Are there any flights to Washington that arrive before noon?" the user interface system may have had trouble during speech recognition, the command may be unparseable given the currently active syntactic grammar, the question may be semantically incomplete (e.g., there are three airports in the Washington, D.C. metropolitan region), or the question may simply be beyond the capability of the interface system or may request information that is not available in the active database. In such cases, the well designed dialogue interface agent will ask the user to repeat or rephrase the question, or will provide more useful indicators of the sort of difficulty it is having in processing the input.
In our wizard study (Miller, et al. 1995) of one type of spoken dialogue user interface system, namely, computer-mediated human-human dialogue, we discovered an unforseen problem having to do with repair dialogues: at the end of a repair dialogue sequence, human participants were uncertain as to how to return to the primary dialogue, and often quite awkward in accomplishing the return. In our prior work on generic conversational dialogue (Duff, et al. 1996; LuperFoy, 1996) we have proposed a four step process for recovering from dialogue trouble.
1. Detection: One of the parties in the dialogue must recognize that there is a problem, otherwise the dialogue continues, possibly hampered by a propogated miscommunication.
2. Diagnosis: A determination must be made as to the source of the dialogue trouble. For automated systems, as with human dialogue interaction, failure can take place at a number of levels. We have reused Clark's (1997) eight levels of presentation acceptance to serve as potential points of interpretation failure, and distinguish different categories of failure based on which component of the dialogue agent system indicates difficulty in carrying out its task: speech recognition, natural language processing, discourse interpretation, domain model compatibility, and so on.
3. Determining a repair plan: Even given a successful detection and diagnosis, user-interface design parameters can influence the preferred method for recovery. For example, upon determining that a user misconception is the source of an illegally stated question or command, the system has the options of (a) correcting the command without bringing it to the user's attention, (b) correcting the command and reporting back to the user the proper formulation of the command, or (c) reporting the problem without correcting it and suggesting that the user reissue the command using the legal formulation.
4. Executing the repair plan: For spoken dialogue systems this repair plan must be executed in collaboration with the other dialogue agent(s). Since human dialogue agents are unpredictable in repair dialogues as they are in primary dialogues, there must be allowances for adjusting to user input that fails to match behavior prompted for.
With our focus on these four steps to recovering from a dialogue trouble situation, we have neglected the crucial fifth step* which we propose here.
5. Closure and return to the primary dialogue: Once the dialogue trouble has been repaired, both system and user must be brought jointly to the understanding that the next utterance is to be interpreted as a return to the primary dialogue. In our discussion we will present multiple methods for achieving this closure step through a separate "closure" statement or through appropriate embellishment of the next utterance to unambiguously associate it with the primary dialogue.
*We acknowledge prevention of dialogue trouble as another important issue for any serious work on repair dialogue processing. This issue has been addressed by McRoy (1996) and others, and is not a focus of our current work.
Duff, D. and S. LuperFoy (1996) "A Centralized Troubleshooting Mechanism for a Spoken Dialogue Interface to a Simulation Application" International Symposium on Spoken Dialogue, Philadelphia.
LuperFoy, S. (1996) "Tutoring Versus Training: A Spoken Language Dialogue Manager for Instructional Systems" TWLT-11 Twente Workshop on Language Technology Number 11, Dialogue Management in Natural Language Systems. University of Twente.
LuperFoy, S. and D. Duff (1996) "Diso: A Four-Step Dialogue Recovery Program" AAAI workshop on Prevention, Detection, and Repair of Dialogue Errors.
McRoy, S. and G. Hirst (1995) "The Repair of Speech Act Misunderstandings by Abductive Inference" Journal of Computational Linguistics, vol. 21 no. 4.
Miller, K., S. LuperFoy, E. Kim, D. Duff (1995) "Some Effects of Electronic Mediation on Spoken Bilingual Dialogue: An Observational Study of Dialogue Management for the Interpreting Telephone" Electronic Journal of Communication.