Conclusions

In this paper, we described an intelligent adaptive conversational assistant designed to help people select an item. Overall, we made significant inroads into methods for unobtrusively acquiring an individual, long term user model during recommendation conversations. We expanded on previous work on adaptive recommendation systems that were not conversational, and on dialogue systems that were not user adaptive. Our long-term goal is to develop even more powerful methods, capable of adapting to a user's needs, goals, and preferences over multiple conversations. While we leveraged off the feedback between conversation and recommendation, such feedback is likely to be present in other tasks such as planning or scheduling. The two key problems addressed by our research are the design of adaptive recommendation systems when conversations are the interaction mode, and the addition of personalization to dialogue systems, starting here with dialogues for recommendation. Thus, unlike many recommendation systems that accept keywords and produce a ranked list, this one carries out a conversation with the user to progressively narrow his options. In solving these problems, we introduced a novel approach to the acquisition, use, and representation of user models. Unlike many other adaptive interfaces, our system constructs and utilizes user models that include information beyond complete item preferences. This is key for the support of personalization in conversations. We used a relatively simple model of dialogue to focus on the issues involved in personalization. We also described experimental results showing the promise of our technique, demonstrating a reduction in both the number of interactions and in the conversation time for users interacting with our adaptive system when compared to a control group. Of course, there are still several open questions and opportunities for improvement. The user model, conversational model, and search models are functional but we plan to improve them further. We are also extending our conversational approach to items other than destinations, such as books and movies, and we plan to link the system to other assistants like the Adaptive Route Advisor [65]. Our goal for such additions is to provide new functionality that will make the ADAPTIVE PLACE ADVISOR more attractive to users, but also to test the generality of our approach for adaptive recommendation. In turn, this should bring us closer to truly flexible computational aides that carry out natural dialogues with humans. Acknowledgments This research was carried out while the first author was at the Center for the Study of Language and Information, Stanford University, and the other authors were at the DaimlerChrysler Research and Technology Center in Palo Alto, California. We thank Renée Elio, Afsaneh Haddadi, and Jeff Shrager for the initial conception and design of the ADAPTIVE PLACE ADVISOR, Cynthia Kuo and Zhao-Ping Tang for help with the implementation effort, and Stanley Peters for enlightening discussions about the design of conversational interfaces. Robert Mertens and Dana Dahlstrom were crucial in carrying out the user studies.