Conversational Interfaces

There is considerable ongoing work in the area of conversational systems, as evidenced in the general surveys by [27] and [54]. [83] give a more thorough overview of user modeling in dialogue systems. [64] reported one of the earliest (typewritten) conversational interfaces, which focused on book recommendation. At the beginning of an interaction, the system asked several questions to place the user in a stereotype group, thereby initializing the user model. As each conversation progressed, this model was adjusted, with the system using ratings to represent uncertainty. However, the language understanding capabilities of the system were limited, mostly allowing only yes/no user answers. More recently, dialogue systems utilize models of user's beliefs and intentions to aid in dialogue management and understanding, though typically these systems maintain models only over the course of a single conversation [46]. As noted in Section 2.3, an important distinction is whether only one conversational participant keeps the initiative, or whether the initiative can switch between participants. Two ambitious mixed-initiative systems for planning tasks are TRAINS [8] and more recent TRIPS [7]. Like the PLACE ADVISOR, these programs interact with the user to progressively construct a solution, though the knowledge structures are partial plans rather than constraints, and the search involves operators for plan modification rather than for database contraction and expansion. TRAINS and TRIPS lack any mechanism for user modeling, but the underlying systems are considerably more mature and have been evaluated extensively. [73] describe another related mixed-initiative system with limited user modeling, in this case a conversational interface for circuit diagnosis. Their system aims to construct not a plan or a set of constraints, but rather a proof tree. The central speech act, which requests knowledge from the user that would aid the proof process, is invoked when the program detects a `missing axiom' that it needs for its reasoning. This heuristic plays the same role in their system as does the PLACE ADVISOR's heuristic for selecting attributes to constrain during item selection. The interface infers user knowledge during the course of only a single conversation, not over the long term as in our approach. With respect to dialogue management, several previous systems have used a method similar to our frame-based search. In particular, Seneff et al. (1998) and Dowding et al. (1993) developed conversational interfaces that give advice about air travel. Like the PLACE ADVISOR, their systems ask the user questions to reduce the number of candidates, treating flight selection as the interactive construction of database queries. However, the question sequence is typically fixed in advance, despite the clear differences among individuals in this domain. Also, these systems usually require that all constraints be specified before item presentation begins. An alternative technique for selecting which questions to ask during information elicitation is presented in [62]. Their overall system necessitates that the system recognize plans the user is attempting to carry out. Then the system must decide how to best complete those plans. When insufficient information is available for plan formation, their system enters an information seeking subdialogue similar to the constraint-satisfaction portion of our dialogues. Their system can decide which question to ask based on domain knowledge or based on the potential informativeness of the question. Another approach to dialogue management is ``conversational case-based reasoning'' (Aha, Breslow & Muñoz-Avila, 2001), which relies on interactions with the user to retrieve cases (items) that will recommend actions to correct some problem. The speech acts and basic flow of control have much in common with the ADAPTIVE PLACE ADVISOR, in that the process of answering questions increasingly constrains available answers. One significant difference is that their approach generates several questions or items, respectively, at a time, and the user selects which question to answer or which item is closest to his or her needs, respectively. Finally, our approach draws on an alternative analysis of item recommendation, described by [30,31]. The main distinctions from that work are that their approach does not include personalization, that they distinguish between search through a task space and through a discourse space, while we combine the two, and that they place a greater emphasis on user intentions. Keeping a distinction between the task and the discourse space in a personalized system would unnecessarily complicate decisions about when to perform user model updates and about how to utilize the model.