Combination of 3 Types of Speech Recognizers for Anaphora Resolution

Combination of 3 Types of Speech Recognizers for Anaphora Resolution � Kazutaka Shimada, Noriko Tanamachi, and Tsutomu Endo Department of Artificial Intelligence, Kyushu Institute of Technology 680-4 Iizuka Fukuoka Japan 820-8502 �shimada, n tanamachi, endo�@pluto.ai.kyutech.ac.jp Abstract. In this paper, we propose a method for anaphora resolution in speech under- standing for a livelihood support robot. For robust speech recognition, we combine two types of speech recognizers; a large vocabulary continuous speech recognizer (LVCSR) and domain-specific speech recognizers (DSSR). One problem in the anaphora resolution is lack of the antecedent in the outputs. To solve the problem, we introduce 2 types of DSSRs; one medium-scale DSSR and several small DSSRs. In this paper, we describe the basic idea of our multiple speech recognizer first. The selection process in the recognizer is based on the similarity between the LVCSR and each DSSR. Then, by using the outputs from the LVCSR and the medium-scale DSSR, we resolve anaphoric expressions in the current output from a small-scale DSSR. The experimental result shows the effectiveness of our method. Keywords: Anaphora resolution, Multiple speech recognizer, Combination 1 Introduction Speech understanding and dialogue systems have been developed for practical use recently. These systems often recognize user utterances incorrectly. It is important to deal with speech recognition errors for speech understanding systems. Extracting keywords and understanding an utterance using them reduce speech recognition errors (Bouwman et al., 1999; Komatani and Kawahara, 2000). Combining some recognizers is one of the best approaches to improve the accuracy of speech understanding systems (Isobe et al., 2007; Utsuro et al., 2004). Utsuro et. al. (2004) have obtained high accuracy by using some speech recognizers’ outputs. However they dealt with word error reduction only. Although Isobe et. al. (2007) have proposed a multi-domain speech recognition system based on some domain-specific recognizers, their system cannot treat out-of- domain utterances such as a chat between users. However chat utterances often include significant information as the context of the dialogue. In this paper we propose a simple and effective speech understanding method based on a large vocabulary continuous speech recognizer (LVCSR) and some domain-specific speech recognizers (DSSR). We call it “One Generalist and Some Specialists (OGSS) model”. Figure 1 (a) shows the outline of the model. In our system, the LVCSR is the generalist, namely domain-independent, and the DSSRs are specialists, namely domain-dependent. We focus on the difference between outputs generated from the generalist and specialists. By using this method, we can recog- nize domain-dependent speech inputs with high accuracy and also handle context information in domain-independent speech inputs. The task of this system is speech understanding for a livelihood support robot. The DSSRs recognize particular utterances about orders; e.g., order utterances from elders who need care and order utterances from nurses. We construct the grammar-based DSSRs for order utterances with � This research was supported by NEDO, Intelligent RT Software Project, 2010. Copyright 2010 by Kazutaka Shimada, Noriko Tanamachi, and Tsutomu Endo PACLIC 24 Proceedings 281 OGSS model (One Generalist and Some Specialists) LVCSR Generalist …

Content truncated to 500 words