A Case-Based Reasoning Approach for Speech Corpus Generation

R. Dale et al. (Eds.): IJCNLP 2005, LNAI 3651, pp. 993 – 1003, 2005. © Springer-Verlag Berlin Heidelberg 2005 A Case-Based Reasoning Approach for Speech Corpus Generation Yandong Fan and Elizabeth Kendall School of Network Computing, Faculty of IT, Monash University, Australia yandong.fan@infotech.monash.edu.au Kendall@infotech.monash.edu.au Abstract. Corpus-based stochastic language models have achieved significant success in speech recognition, but construction of a corpus pertaining to a specific application is a difficult task. This paper introduces a Case-Based Reasoning system to generate natural language corpora. In comparison to traditional natural language generation approaches, this system overcomes the inflexibility of template-based methods while avoiding the linguistic sophistication of rule-based packages. The evaluation of the system indicates our approach is effective in generating users’ specifications or queries as 98% of the generated sentences are grammatically correct. The study result also shows that the language model derived from the generated corpus can significantly outperform a general language model or a dictation grammar. 1 Introduction Stochastic language models have achieved significant success in speech recognition since the last decade [1, 2]. The main underlying technique in stochastic approaches is the use of corpora. The successful utilization of corpora has been proven by many researchers [3,4,5]. However, construction of a corpus pertaining to a specific application is a difficult task, given that there is no pre-knowledge on how users might communicate with the application before the deployment of a system. Research [6] has been conducted to explore the effectiveness of Web-based text corpus generation. Although the proliferation of eText has made the collection of textual material easier than ever, Thompson [7] argues that actually locating eText appropriate to your needs can be quite difficult. Moreover, in the context of conversational systems, the suitability of corpora purely collected from the Internet is controversial due to the difference of written text and spoken language. Although generally there is a lack of spoken material pertaining to a new application, ample transcriptions do exist in some well-established domains, such as the Air Traffic Information System (ATIS) domain. User modeling has been studied for long and conversations between users and agents of spoken language systems have been recorded and accumulated for decades in these domains. In our project, we seek to develop a speech-enabled mobile commerce application, which we called the MCCS (Mobile Car City system). The system allows a mobile-phone user to specify preferences or initiate queries by speech at the beginning of the conversation. Then all 994 Y. Fan and E. Kendall car models conforming to the preferences or queries are retrieved to guide the user in finding specific car models that meet his/her needs. Through carefully examining the spoken transcriptions from the ATIS domain, we believe that user specifications or queries in the MCCS system should share significant similarity in sentence syntactic structure with their counterparts in the ATIS domain. Motivated by this assumption, we believe that the MCCS system can learn a set of sample sentences from the ATIS domain, which can then be used as the knowledge base …

Content truncated to 500 words