Answering Complex Questions via Exploiting Social Q&A Collection

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 956–964, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Answering Complex Questions via Exploiting Social Q&A Collection Youzheng Wu Chiori Hori Hisashi Kawai Hideki Kashioka Spoken Language Communication Group, MASTAR Project National Institute of Information and Communications Technology (NiCT) 2-2-2 Hikaridai, Keihanna Science City, Kyoto 619-0288, Japan {youzheng.wu, chiori.hori, hisashi.kawai, hideki.kashioka}@nict.go.jp Abstract This paper regards social Q&A collec- tions, such as Yahoo! Answer as a knowl- edge repository and investigates tech- niques to mine knowledge from them for improving a sentence-based complex question answering (QA) system. In par- ticular, we present a question-type-specific method (QTSM) that studies at extract- ing question-type-dependent cue expres- sions from the social Q&A pairs in which question types are the same as the submitted question. The QTSM is also compared with question-specific and monolingual translation-based meth- ods presented in previous work. There- into, the question-specific method (QSM) aims at extracting question-dependent an- swer words from social Q&A pairs in which questions are similar to the submit- ted question. The monolingual translation- based method (MTM) learns word-to- word translation probabilities from all so- cial Q&A pairs without consideration of question and question type. Experiments on extension of the NTCIR 2008 Chinese test data set verify the performance rank- ing of these methods as: QTSM > {QSM, MTM}. The largest F3 improvements of the proposed QTSM over the QSM and MTM reach 6.0% and 5.8%, respectively. 1 Introduction Research on the topic of QA systems has mainly concentrated on answering factoid, definitional, reason and opinion questions. Among the ap- proaches proposed for answering these questions, machine learning techniques have been found more effective in constructing QA components from scratch. Yet these supervised techniques require a certain scale of question and answer (Q&A) pairs as training data. For example, Echi- habi et al. (2003) and Sasaki (2005) respectively constructed 90,000 English and 2,000 Japanese Q&A pairs for their factoid QA systems. Cui et al. (2004) collected 76 term-definition pairs for their definitional QA system. Higashinaka and Isozaki (2008) used 4,849 positive and 521,177 negative examples in their reason QA system. Stoyanov et al. (2005) required a known subjective vocabulary for their opinion QA system. This paper is con- cerned with answering complex questions which answers generally consists of a list of nuggets (Voorhees, 2003; Mitamura et al., 2008). Apart from definitional and opinion (TAC, 2008) com- plex questions, many other types of complex ques- tions have not yet to be thoroughly studied1. To answer these complex questions via supervised techniques, we need to collect training Q&A pairs for each type of complex question, though this is an extremely expensive and labor-intensive task. This paper is to explore the possibility of auto- matic learning of training Q&A pairs and mining needed knowledge from social Q&A collections such as Yahoo! Answer2. That is to say, we are interested in whether or not millions of, possible noisy, user-generated Q&A pairs can be exploited for automatic QA system. …

Content truncated to 500 words