An Online Algorithm for Applying Reinforcement Learning to Handle Ambiguity in Spoken Dialogues

Abstract

Spoken dialogue systems (SDSs) have been widely used in human-computer communications, including database querying, online trouble shooting advising, etc. A major challenge in building an SDS is to handle ambiguity in natural languages. User queries, questions, descriptions in a natural language may be ambiguous. To be effective in practical applications, an SDS must be able to disambiguate input from its user(s). In our research, we develop an online algorithm for applying reinforcement learning to handle ambiguity in SDSs. We introduce a new user dialogue policy into the framework of reinforcement learning to model user dialogue behavior. Also, differing from the current reinforcement learning algorithms in speech and language processing that are characterized by offline training, our algorithm conducts both offline and online detection of user dialogue behavior. In this paper, we present the online algorithm for reinforcement learning, emphasizing the detection of user dialogue behavior. We also describe the initial implementation and experiments.