Introduction and Motivation

Recommendation systems help users find and select items (e.g., books, movies, restaurants) from the huge number available on the web or in other electronic information sources [16,63,17]. Given a large set of items and a description of the user's needs, they present to the user a small set of the items that are well suited to the description. Recent work in recommendation systems includes intelligent aides for filtering and choosing web sites [29], news stories [11], TV listings [24], and other information. The users of such systems often have diverse, conflicting needs. Differences in personal preferences, social and educational backgrounds, and private or professional interests are pervasive. As a result, it seems desirable to have personalized intelligent systems that process, filter, and display available information in a manner that suits each individual using them. The need for personalization has led to the development of systems that adapt themselves by changing their behavior based on the inferred characteristics of the user interacting with them [10,32,33,50,64]. The ability of computers to converse with users in natural language would arguably increase their usefulness and flexibility even further. Research in practical dialogue systems, while still in its infancy, has matured tremendously in recent years [7,27,54]. Today's dialogue systems typically focus on helping users complete a specific task, such as planning, information search, event management, or diagnosis. In this paper, we describe a personalized conversational recommendation system designed to help users choose an item from a large set all of the same basic type. Our goal is to support conversations that become more efficient for individual users over time. Our system, the ADAPTIVE PLACE ADVISOR, aims to help users select a destination (in this case, restaurants) that meets their preferences. The ADAPTIVE PLACE ADVISOR makes three novel contributions. To our knowledge, this is the first personalized spoken dialogue system for recommendation, and one of the only conversational natural language interfaces that includes a personalized, long-term user model. Second, it introduces a novel model for acquiring, utilizing, and representing user models. Third, it is used to demonstrate a reduction in the number of system-user interactions and the conversation time needed to find a satisfactory item. The combination of dialogue systems with personalized recommendation addresses weaknesses of both approaches. Most dialogue systems react similarly for each user interacting with them, and do not store information gained in one conversation for use in the future. Thus, interactions tend to be tedious and repetitive. By adding a personalized, long-term user model, the quality of these interactions can improve drastically. At the same time, collecting user preferences in recommendation systems often requires form filling or other explicit statements of preferences on the user's part, which can be difficult and time consuming. Collecting preferences in the course of the dialogue lets the user begin the task of item search immediately. The interaction between conversation and personalized recommendation has also affected our choices for the acquisition, utilization, and representation of user models. The ADAPTIVE PLACE ADVISOR learns information about users unobtrusively, in the course of a normal conversation whose purpose is to find a satisfactory item. The system stores this information for use in future conversations with the same individual. Both acquisition and utilization occur not only when items are presented to and chosen by the user, but also during the search for those items. Finally, the system's representation of models goes beyond item preferences to include preferences about both item characteristics and particular values of those characteristics. We believe that these ideas extend to other types of preferences and other types of conversations. In this paper, we describe our work with the ADAPTIVE PLACE ADVISOR. We begin by introducing personalized and conversational recommendation systems, presenting our design decisions along the way. In Section 3 we describe the system in detail, while in Section 4 we present our experimental evaluation. In Sections 5 and 6 we discuss related and future work, respectively, and in Section 7 we conclude and summarize the paper.