Adaptive Dialogue Systems

Finally, another body of recent work describes the use of machine learning or other forms of adaptation to improve dialogue systems.¹⁰ Researchers in this area develop systems that learn user preferences, improve task completion, or adapt dialogue strategies to an individual during a conversation. The closest such work also pursues our goal of learning user preferences. [19] report one such example for consultation dialogues, but take a different approach. Their system acquires value preferences by analyzing both user's explicit statements of preferences and their acceptance or rejection of the system's proposals. It uses discrete preference values instead of our more fine-grained probability model. Also, their system does not use preferences during item search but only at item presentation time to help evaluate whether better alternatives exist. Finally, their evaluation is based on subject's judgements of the quality of the system's hypotheses and recommendations, not on characteristics of actual user interactions. We could, however, incorporate some of their item search ideas, allowing near misses between user-specified constraints and actual items. Another system that focuses on user preferences is an interactive travel assistant [52] that carries out conversations via a graphical interface. The system asks questions with the goal of narrowing down the available candidates, using speech acts similar to ours, and also aims to satisfy the user with as few interactions as possible. Their approach to minimizing the number of interactions is to use a candidate/critique approach. From a user's responses, the system infers a model represented as weights on attributes such as price and travel time. Unlike the ADAPTIVE PLACE ADVISOR, it does not carry these profiles over to future conversations, but one can envision a version that does so. Several authors use reinforcement learning techniques to improve the probability of or process of task completion in a conversation. For example, Singh et al. (2002) use this approach to determine the system's level of initiative and the amount of confirmation of user utterances. Their goal is to optimize, over all users, the percentage of dialogues for which a given task is successfully completed. This system leverages the learned information when interacting with all users, rather than personalizing the information. Also, [51] use reinforcement learning to determine which question to ask at each point during an information seeking search, but do not demonstrate the utility of their approach with real users. Finally, a number of systems adapt their dialogue management strategy over the course of a conversation based on user responses or other dialogue characteristics. For example, [53] use a set of learned rules to decide whether a user is having difficulty achieving their task, and modify the level of system initiative and confirmation accordingly. [55] present a help-desk application that first classifies the user as a novice, moderate, or expert based on responses to prompts. It then adjusts the complexity of system utterances, the jargon, and the complexity of the path taken to achieve goals. [39] apply user modeling to a dialogue system that uses evidence from the current context and conversation to update a Bayesian network. The network influences the spoken language recognition hypothesis and causes appropriate adjustments in the system's level of initiative. [21] describes a system that adapts both language generation and initiative strategies for an individual user within a single dialogue. Also, Jameson et al. (1994) use Bayesian networks in a system that can take the role of either the buyer or seller in a transaction, and that changes its inquiry or sales strategy based on beliefs inferred from the other participant's utterances.