Chinese Speech Information Retrieval for Questions on Mobile Phone Operation

Chinese Speech Information Retrieval for Questions on Mobile Phone Operation Kai Ishikawa1, Susumu Akamine1, and Ken Hanazawa1 1 Media and Information Research Laboratories, NEC Corporation 1753, Shimonumabe, Nakahara-ku, Kawasaki, Kanagawa 211-8666, Japan {k-ishikawa@dq, s-akamine@ak, k-hanazawa@cq}.jp.nec.com Abstract. We propose an information retrieval method based on question identification. We have implemented and integrated a prototype Chinese speech IR system for a commercial PDA with this method for questions concerning mobile phone operations. A preliminary evaluation showed that our method is effective. In the evaluation, our method outperformed a similar document search method by ten points or more in both its precision and recall for top-ranked candidates. Keywords: Information Retrieval, Chinese Speech Recognition, Question Identification, Question and Answering, FAQ, User’s Manual. 1 Introduction As IT devices are becoming more complex and multifunctional, users are finding it hard to learn how to properly use all the available functions. Information retrieval is a useful technology for users to find the answers to their questions at anytime. Ikeda et al. applied a conventional similar document search function to a user’s manual document (Ikeda et al. 2004). However, compared to the Web or other open domain retrieval tasks, a manual/FAQ search is a very domain-specific task, where the terms used in documents and queries are limited, and this causes high overlapping among the terms of different queries. A limited domain would adversely influence the similar document search function, because it evaluates the score of a document by counting the query terms in the document. To overcome this problem, Lytinen et al. proposed a FAQ retrieval method with question types, which are currently used by several question-answering systems competing in the TREC Q&A track (Abney et al. 2000, Cardie et al. 2000, and Harabagiu et al. 2000), and this method has shown improvements in the performance (Lytinen et al. 2002). This method can improve the precision of FAQ candidates by taking into consideration the similarities in a query’s question type and a FAQ’s question type. However, a question type is not a sufficiently detailed and reliable information to be used in the manual/FAQ retrieval system on a mobile device. Therefore, our idea is to solve the manual/FAQ retrieval problem by using an identification approach along with a conventional retrieval approach. We propose in this paper a hybrid information retrieval method combined with a request identifier based on a syntactic parser. The parser uses “generic” syntactic rules, while requesting “specific” syntactic rules at the same time, and outputs the most plausible request ID according to the disambiguation of the syntactic structure. By introducing the identifier of a user’s request within a particular domain, the method can provide highly precise answers. We collected 10,000 request sentences from a mobile phone operation domain, clustered the sentences into groups, and manually assigned a unique request ID to each set. An answer document (or a few answers) for each ID was selected from the corresponding user’s manual documents. In our method, we first introduce a request identifier based on syntactic parsing to identify …

Content truncated to 500 words