Efficient Confirmation Strategy for Large-scale Text Retrieval Systems with Spoken Dialogue Interface

Efficient Confirmation Strategy for Large-scale Text Retrieval Systems with Spoken Dialogue Interface Kazunori Komatani Teruhisa Misu Tatsuya Kawahara Hiroshi G. Okuno Graduate School of Informatics Kyoto University Yoshida-Hommachi, Sakyo, Kyoto 606-8501, Japan {komatani,kawahara,okuno}@i.kyoto-u.ac.jp Abstract Adequate confirmation for keywords is in- dispensable in spoken dialogue systems to eliminate misunderstandings caused by speech recognition errors. Spoken lan- guage also inherently includes out-of- domain phrases and redundant expressions such as disfluency, which do not contribute to task achievement. It is necessary to appropriately make confirmation for im- portant portions. However, a set of key- words necessary to achieve the tasks can- not be predefined in retrieval for a large- scale knowledge base unlike conventional database query tasks. In this paper, we describe two statistical measures for iden- tifying portions to be confirmed. A rele- vance score represents the matching degree with the target knowledge base. A sig- nificance score detects portions that conse- quently affect the retrieval results. These measures are defined based on information that is automatically derived from the tar- get knowledge base. An experimental eval- uation shows that our method improved the success rate of retrieval by generating con- firmation more efficiently than using a con- ventional confidence measure. 1 Introduction Information retrieval systems with spoken lan- guage have been studied (Harabagiu et al., 2002; Hori et al., 2003). They require both automatic speech recognition (ASR) and information re- trieval (IR) technologies. As a straight mani- festation to create these systems, we can think of using ASR results as an input for IR systems that retrieve a text knowledge base (KB). How- ever, two problems occur in the characteristics of speech. 1. Speech recognition errors 2. Redundancy included in spoken language expressions One is an ASR error, which is basically in- evitable in speech communications. Therefore, an adequate confirmation is indispensable in spoken dialogue systems to eliminate the mis- understandings caused by ASR errors. If keywords to be confirmed are defined, the system can confirm them using confidence mea- sures (Komatani and Kawahara, 2000; Hazen et al., 2000) to manage the errors. In con- ventional tasks for spoken dialogue systems in which their target of retrieval was well-defined, such as the relational database, keywords that are important to achieve the tasks correspond to items in the relational database. Most spo- ken dialogue systems that have been developed, such as airline information systems (Levin et al., 2000; Potamianos et al., 2000; San-Segundo et al., 2000) and train information systems (Allen et al., 1996; Sturm et al., 1999; Lamel et al., 1999), are categorized here. However, it is not feasible to define such keywords in retrieval for operation manuals (Komatani et al., 2002) or WWW pages, where the target of retrieval is not organized and is written as natural language text. Another problem is that a user’s utterances may include redundant expressions or out-of- domain phrases. A speech interface has been said to have the advantage of ease of input. This means that redundant expressions, such as dis- fluency and irrelevant phrases, …

Content truncated to 500 words