Speech-based Information Retrieval System with Clarification Dialogue Strategy

Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP), pages 1003–1010, Vancouver, October 2005. c⃝2005 Association for Computational Linguistics Speech-based Information Retrieval System with Clarification Dialogue Strategy Teruhisa Misu Tatsuya Kawahara School of informatics Kyoto University Sakyo-ku, Kyoto, Japan misu@ar.media.kyoto-u.ac.jp Abstract This paper addresses a dialogue strategy to clarify and constrain the queries for speech-driven document retrieval systems. In spoken dialogue interfaces, users often make utterances before the query is com- pletely generated in their mind; thus input queries are often vague or fragmental. As a result, usually many items are matched. We propose an efficient dialogue frame- work, where the system dynamically se- lects an optimal question based on infor- mation gain (IG), which represents reduc- tion of matched items. A set of possible questions is prepared using various knowl- edge sources. As a bottom-up knowl- edge source, we extract a list of words that can take a number of objects and po- tentially causes ambiguity, using a depen- dency structure analysis of the document texts. This is complemented by top-down knowledge sources of metadata and hand- crafted questions. An experimental evalu- ation showed that the method significantly improved the success rate of retrieval, and all categories of the prepared questions contributed to the improvement. 1 Introduction The target of spoken dialogue systems is being ex- tended from simple databases such as flight informa- tion (Levin et al., 2000; Potamianos et al., 2000) to general documents (Fujii and Itou, 2003) including newspaper articles (Chang et al., 2002; Hori et al., 2003). In such systems, the automatic speech recog- nition (ASR) result of the user utterance is matched against a set of target documents using the vector space model, and documents with high matching scores are presented to the user. In this kind of document retrieval systems, user queries must include sufficient information to iden- tify the desired documents. In conventional doc- ument query tasks with typed-text input, such as TREC QA Track (NIST and DARPA, 2003), queries are (supposed to be) definite and specific. However, this is not the case when speech input is adopted. The speech interface makes input easier. However, this also means that users can start utterances before queries are thoroughly formed in their mind. There- fore, input queries are often vague or fragmental, and sentences may be ill-formed or ungrammatical. Moreover, important information may be lost due to ASR errors. In such cases, an enormous list of possi- ble relevant documents is usually obtained because there is very limited information that can be used as clues for retrieval. Therefore, it is necessary to narrow down the documents by clarifying the user’s intention through a dialogue. There have been several studies on the follow-up dialogue, and most of these studies assume that the target knowledge base has a well-defined structure. For example, Denecke (Denecke and Waibel, 1997) addressed a method to generate guiding questions based on a tree structure constructed by unifying pre-defined keywords and semantic slots. However, these …

Content truncated to 500 words