Language Modeling for Spoken Dialogue System based on Filtering using Predicate-Argument Structures

Proceedings of COLING 2012: Technical Papers, pages 2993–3002, COLING 2012, Mumbai, December 2012. Language Modeling for Spoken Dialogue System based on Filtering using Predicate-Argument Structures Koichiro Yoshino1 Shinsuke Mori1 Tatsuya Kawahara1 (1) School of Informatics, Kyoto University, Yoshida-Honmachi, Sakyo-ku, Kyoto, Japan. ABSTRACT We present a novel scheme of language modeling for a spoken dialogue system by effectively filtering query sentences collected via a Web site of wisdom of crowds. Our goal is a speech- based information navigation system by retrieving from backend documents such as Web news. Then, we expect that users make queries that are relevant to the backend documents. The relevance measure can be defined with cross-entropy or perplexity by the language model generated from the documents in a conventional manner. In this article, we propose a novel criteria that considers semantic-level information. It is based on predicate-argument (P-A) pairs and their relevance to the documents (or topic) is defined by a naive Bayes score. Experimental evaluations demonstrate that the proposed relevance measure effectively selects relevant sentences used for a language model, resulting in significant reduction of the word error rate of speech recognition as well as the semantic-level error rate. KEYWORDS: Language Modeling, Predicate Argument Structure, Spoken Dialogue System. 2993 1 Introduction The tasks of spoken dialogue systems have been extended from simple transactions to general information navigation based upon user requests. Ideally, these systems should handle not only simple, keyword-based queries that current voice search systems respond to but also vague and complex user requests related to, for example, tourist guides or news briefings. This type of application can be achieved through document retrieval in a corresponding domain. For example, we can turn to tourist guidebooks or relevant Wikipedia entries for information on the tourist domain (Misu and Kawahara, 2010). An intelligent dialogue system can be created by restricting the domain and using the knowledge from that domain (Kawahara, 2009). An interactive news navigator that generates dialogues based on news article archives has been developed along this concept (Yoshino et al., 2011). The automatic speech recognition (ASR) module for spoken dialogue systems (SDSs) needs an appropriate language model (LM) adapted to the task domain and style. Even an ASR system with a very large vocabulary cannot cover all proper nouns or named entities (NEs), which are critical in information retrieval. Ideally, an LM should be trained with a large-scale matched corpus, but in many cases this is not realistic. Therefore, two approaches are commonly adopted. The first involves mixing document texts of the target domain with a dialogue corpus of spoken-style expressions. The other involves collecting relevant texts, possibly from spoken- style sentences, from the Web (Sarikaya et al., 2005; Sethy et al., 2005; Misu and Kawahara, 2006; Bulyko et al., 2007). These approaches try to cover the target domain and style of speech in an indirect way, but the resultant model will inevitably contain a large amount of irrelevant texts. In general, information navigation systems with speech interfaces have a set of backend documents for …

Content truncated to 500 words