A Spoken Dialogue System for Document Information Retrieval Utilizing Topic Knowledge

摘要

The authors implemented a mechanism for using retrieval field information to promote efficient searching in a spoken dialogue system for document information retrieval tasks. The research focused mainly on the following two points: (1) correctly categorizing the documents in the database into target topics and (2) properly using topic knowledge to promote efficient searching. For point (1), the authors proposed a technique for recursively calculating the relevance score between (retrieval) words and topics (RSW) and the relevance score between documents and topics (RSD). A maximum topic categorization rate of 86.1% was obtained as a result of experimental investigations. An improvement of 5.0% from the initial value (81.1 %) due to the recursive calculations showed the effectiveness of the proposed technique. Subjective evaluations were also used to verify the effectiveness of the relevance scores obtained by the proposed technique during practical use. For point (2), the authors compared the proposed system with a system that does not use topic knowledge when the retrieval field was estimated to advance the dialogue. It was apparent that the proposed system could decrease the number of retrieval dialogue interactions and the dialogue interaction period by approximately 20%, which showed the effectiveness of using topic knowledge for aiding efficient searching.