Action State Update Approach to Dialogue Management

摘要: Utterance interpretation is one of the main functions a dialogue manager, which key component system. We propose action state update approach (ASU) for utterance interpretation, featuring statistically trained binary classifier used to detect actions in text user utterance. Our goal interpret referring expressions input without domain-specific natural language understanding component. For training model, we use active learning automatically select simulated examples. With both user-simulated and interactive human evaluations, show that ASU successfully interprets utterances system, including those with expressions.