Corpus-based Question Answering for why-Questions

Corpus-based Question Answering for why-Questions Ryuichiro Higashinaka and Hideki Isozaki NTT Communication Science Laboratories, NTT Corporation 2-4, Hikaridai, Seika-cho, Kyoto 619-0237, Japan {rh,isozaki}@cslab.kecl.ntt.co.jp Abstract This paper proposes a corpus-based ap- proach for answering why-questions. Con- ventional systems use hand-crafted patterns to extract and evaluate answer candidates. However, such hand-crafted patterns are likely to have low coverage of causal expres- sions, and it is also difficult to assign suit- able weights to the patterns by hand. In our approach, causal expressions are automati- cally collected from corpora tagged with se- mantic relations. From the collected expres- sions, features are created to train an an- swer candidate ranker that maximizes the QA performance with regards to the corpus of why-questions and answers. NAZEQA, a Japanese why-QA system based on our ap- proach, clearly outperforms a baseline that uses hand-crafted patterns with a Mean Re- ciprocal Rank (top-5) of 0.305, making it presumably the best-performing fully imple- mented why-QA system. 1 Introduction Following the trend of non-factoid QA, we are seeing the emergence of work on why-QA; e.g., answering generic “why X?” questions (Verberne, 2006). However, since why-QA is an inherently dif- ficult problem, there have only been a small number of fully implemented systems dedicated to solving it. Recent systems at NTCIR-61 Question Answer- ing Challenge (QAC-4) can handle why-questions (Fukumoto et al., 2007). However, their perfor- mance is much lower (Mori et al., 2007) than that of factoid QA systems (Fukumoto et al., 2004; Voorhees and Dang, 2005). We consider that this low performance is due to the great amount of hand-crafting involved in the 1http://research.nii.ac.jp/ntcir/ntcir-ws6/ws-en.html systems. Currently, most of the systems rely on hand-crafted patterns to extract and evaluate answer candidates (Fukumoto et al., 2007). Such patterns include typical cue phrases and POS-tag sequences related to causality, such as “because of” and “by reason of.” However, as noted in (Inui and Okumura, 2005), causes are expressed in various forms, and it is difficult to cover all such expressions by hand. Hand-crafting is also very costly. Some patterns may be more indicativeof causes than others. There- fore, it may be useful to assign different weights to the patterns for better answer candidate extraction, but currently this must be done by hand (Mori et al., 2007). It is not clear whether the weights determined by hand are suitable. In this paper, we propose a corpus-based approach for why-QA in order to reduce this hand-crafting effort. We automatically collect causal expressions from corpora to improve the coverage of causal ex- pressions, and utilize a machine learning technique to train a ranker of answer candidates on the ba- sis of features created from the expressions together with other possible features related to causality. The ranker is trained to maximize the QA performance with regards to a corpus of why-questions and an- swers, automatically tuning the weights of the fea- tures. This paper is organized as follows: Section 2 de- scribes previous work on why-QA, and Section 3 de- scribes our approach. Section 4 describes …

Content truncated to 500 words