GUS, A Frame-Driven Dialog System

GUS is the first o f a series o f experimental computer systems that we intend to construct as part o f a program of research on language understanding. In large measure, these systems will fill the role o f periodic progress reports, summarizing what we have learned, assessing the mutual coherence o f the various lines o f investigation we have been following, and saggestin# where more emphasis is needed in future work. GUS (Genial Understander System) is intended to engage a sympathetic and highly cooperative human in an English dialog, directed towards a specific goal within a very restricted domain o f discourse. As a starting point, G US was restricted to the role o f a travel agent in a conversation with a client who wants to make a simple return trip to a single city in California. There is good reason for restricting the domain o f discourse for a computer system which is to engage in an English dialog. Specializing the subject matter that the system can talk about permiis it to achieve some measure o f realism without encompassing all the possibilities o f human knowledge or o f the English language. It also provides the user with specific motivation for participating in the conversation, thus narrowing the range o f expectations that GUS must have about the user's purposes. A system restricted in this way will be more able to guide the conversation within the boundaries o f its competence. 1. Motivation and Design Issues Within its limitations, ous is able to conduct a more-or-less realistic dialog. But the outward behavior of this first system is not what makes it interesting or significant. There are, after all, much more convenient ways to plan a trip and, unlike some other artificial intelligence programs, (;us does not offer services or furnish information that are otherwise difficult or impossible to obtain. The system is i nteresting because of the phenomena of natural dialog that it attempts to model tThis work was done by the language understander project at the Xerox Palo Alto Research center. Additional affiliations: D. A. Norman, University of California, San Diego; H. Thompso6, University of California, Berkeley; and T. Winograd, Stanford University. Artificial Intelligence 8 0977), 155-173 Copyright © 1977 by North-Holland Publishing Company 156 D . G . BOBI~OW ET AL. and because of the principles of program organization around which it was de, Signed. Among the hallmarks of natural dialogs are unexpected and seemingly unpredictable sequences of events. We describe some of the forms that these can take below. "We then go on to discuss the modular design which makes the system re!atively insensitive t o the vagaries of ordinary conversation. 1.1. Problems of natural dialog The simple dialog shown in Fig. 1 illustrates some of the language-understanding problems we attacked. (The parenthesized numbers are for reference in the text). The problems illustrated in this figure, and described in the paragraphs below, include allowing both the client and the system to take the initiative, understanding indirect answers to questions, resolving anaphora, understanding fragments of sentences offered as answers to questions, and interpreting the discourse in the light of known conversational patterns. 1.1.1. Mixed initiative A typical contribution to a dialog, in addition to its more obvious functions, conveys an expectation about how the other participant will respond. This is clearest in the ease of a question, but it is true of all dialog. If one of the participants has very particular expectations and states them strongly whenever he speaks, and ff the other always responds in such a way as to meet the expectations conveyed, then the initiative remains with the first participant throughout. The success of interactive computer systems can often be traced to the skill with which their designers were able to assure them such a dominating position in the interaction. In natural conversations between humans, however, each participant usually assumes the initiative from time to time. Either clear expectations are not stated or simply not honored. GUS attempts to retain the initiative, but not to the extent of jeopardizing the natural flow of the conversation. To this extent it is a mixed-initiative system (see Carbonell [5, 6]). This is exemplified in the dialogue at (1) where the client volunteers more information than GUS requested. In addition to his destination, the client gives the date on which he wants to travel. Line (3) illustrates a ease where the client takes control of the conversation. GUS had found a potentially acceptable flight and asked for the client's approval. Instead of either giving or denying it, the client replied with a question of his own. 1.1.2. Indirect answers It is by no means always clear what constitutes an answer to a question. Frequently the purported answer is at best only a basis on which to infer the information requested. For example, when ous asks "Whatt ime do you want to leave?" it is seeking information to constrain the selection of a flight, client's res onse t o • • P " . this question, a t (2), does constrain the flight selection, b u t only indirectly. In Artificial Intelligence8 (1977), 155--17a : -