Similarity Based Language Model Construction for Voice Activated Open-Domain Question Answering

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 536–544, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Similarity Based Language Model Construction for Voice Activated Open-Domain Question Answering Istv´an Varga Kiyonori Ohtake Kentaro Torisawa Stijn De Saeger Teruhisa Misu Shigeki Matsuda Jun’ichi Kazama Institute of Information and Communications Technology (NICT) 3-5 Hikaridai, Seika-cho, Soraku-gun, Kyoto, 619-0289 Japan {istvan, kiyonori.ohtake, torisawa, stijn, teruhisa.misu, shigeki.matsuda, kazama}@nict.go.jp Abstract This paper describes a novel method of constructing a language model for speech recognition of inputs with a particular style, using a large-scale Web archive. Our target is an open domain voice-activated QA system and our speech recognition module must recognize relatively short, domain independent questions. The cen- tral issue is how to prepare a large scale training corpus with low cost, and we tack- led this problem by combining an exist- ing domain adaptation method and distri- butional word similarity. From 500 seed sentences and 600 million Web pages we constructed a language model covering 413,000 words. We achieved an average improvement of 3.25 points in word er- ror rate over a baseline model constructed from randomly sampled Web sentences. 1 Introduction This paper presents a novel language model con- struction method for speech recognition, which is to be utilized by Ikkyu, an open-domain voice- activated Japanese QA system. Ikkyu takes rela- tively short spoken questions concerning a broad range of topics as input through a smartphone and provides the answers retrieved from a large scale Web archive. The challenge we tackle here is to provide a language model for a large- vocabulary speech recognition module to recog- nize such questions. The widespread practice in language model construction is to train the model with a corpus that matches the domain and style of the target ap- plication. “Domain” usually refers to a set of utter- ances that are strongly related to a particular topic Figure 1: Screenshot of our QA system (Ikkyu) with the answers for “What causes deflation?”. (e.g., travel). Ikkyu does not have any domain in this sense. The final QA system is expected to an- swer all the questions concerning every possible topic as far as our QA module can find the answers from the Web. Since current state-of-the-art speech recogni- tion systems cannot recognize long sentences through a smartphone with a high sentence accu- racy, we decided to focus on relatively short ques- tions, roughly consisting of a noun, an interroga- tive pronoun, and a predicate. Also, our current QA module can deal with only relatively short questions and this restricts the answerable ques- tions. We call the restrictions on possible ques- tions due to these factors style hereafter. (See Sec- tion 2 for a detailed description of style.) Our challenge is to prepare a large number of questions over various topics that match this style. Manual construction of such a corpus is impossible con- sidering the necessary vocabulary coverage, thus an automatic method for collecting questions is needed. Our method starts from an …

Content truncated to 500 words