Impact of responsive and directive adaptation on local dialog processing

Spoken dialog systems allow users to access information and accomplish tasks using speech. Understanding and interpreting complex and ambiguous natural language phrases is a challenging task for these systems. Adaptation (the phenomenon of one conversational partner’s behavior causing changes in the behavior of the other conversational partner) can be a powerful tool to improve dialog system performance. In this work I examine communication errors in human-computer dialog. I explore the role of directive adaptation (in which the dialog system’s behavior guides the user’s behavior) and responsive adaptation (in which the user’s behavior affects the system’s behavior) in avoiding and fixing these errors. The goal of this work is not to model human interaction, but to design methods for improving dialog systems informed by the model of human communication. The contributions of this thesis include (1) a computational analysis of adaptation in dialog, (2) experiments evaluating user adaptation to the form of system prompts, and (3) experiments evaluating the effect of system adaptation to the content of user utterances. In the first study, I compare two possible explanations for adaptation in dialog: partner design and recency. I propose a new measure of adaptation and use it in a study of the Communicator human-computer spoken dialog corpus to compare strength of adaptation due to recency and to partner design. In the second set of studies, I examine user adaptation to the system’s lexical and syntactic choices in the context of the deployed Let’s Go! dialog system. I show that in deployed dialog systems with real users, as in laboratory experiments, users adapt to the system’s lexical and syntactic choices. I also show that system prompt formulation can be used to guide users into producing utterances conducive to task success. In the third set of studies, I evaluate the effect on speech recognition performance of language model adaptation to the task-related topic and content of user utterances. I show that lexical and dialog history features are useful in prediction of utterance content and that the prior knowledge of the content of a user utterance can lead to improvements in speech recognition performance.