Answering Yes/No Questions via Question Inversion

Proceedings of COLING 2012: Technical Papers, pages 1377–1392, COLING 2012, Mumbai, December 2012. Answering Yes/No Questions via Question Inversion Hiroshi Kanayama1 Y usuke Miyao2 John Prager3 (1) IBM Research - Tokyo, Koto-ku, Tokyo, Japan (2) National Institute of Informatics, Chiyoda-ku, Tokyo, Japan (3) IBM T.J.Watson Research Center, Yorktown Heights, NY, USA hkana@jp.ibm.com, yusuke@nii.ac.jp, jprager@us.ibm.com Abstract This paper investigates a solution to yes/no question answering, which can be mapped to the task of determining the correctness of a given proposition. Generally it is hard to obtain explicit evidence to conclude a proposition is false from an information source, so we convert this task to a set of factoid-style questions and use an existing question answering system as a subsystem. By aggregating the answers and confidence values from a factoid-style question answering system we can determine the correctness of the entire proposition or the substitutions that make the proposition false. We evaluated the system on multiple-choice questions from a university admission test on world history, and found it to be highly accurate. Keywords: question answering, facts validation, yes/no question, question inversion. 1377 1 Introduction Yes/no question answering (Green and Carberry, 1994; Hirschberg, 1984) can be equated to the task of determining the correctness of a given proposition. The target of such a mechanism is not limited to explicit interrogative questions since any general declarative sentences can in principle be validated. For example, consider the following two propositions, where (1) is true and (2) false1: (1) Chirac was the president of France in 2000. ∗(2) Chirac was the president of Germany in 2000. As suggested by this example, a false proposition can often be produced by replacing a part of a true proposition. During a conversation, a human with knowledge of the facts might not only say that the utterance (2) is wrong, but respond with something like “not Germany, but France.” Therefore we believe this kind of validation system we are proposing here may have application outside of a strict question-answering framework. In spite of the importance of the process, yes/no-style question answering has not been intensively studied, compared to other type of question answering, such as the factoid-style question answering (Ravichandran and Hovy, 2002; Bian et al., 2008) and definition question answering (Xu et al., 2003; Cui et al., 2005). Even though there are only two possible answers, yes or no, such questions can be quite hard to answer. Search technologies cannot be applied easily because in many cases we can not rely on the existence of explicit negative evidence in the information source, e.g. “Chirac was not a president of Germany.” This paper tackles yes/no question answering by exploiting an existing system for factoid- style question answering. The key idea, inspired by question inversion (Prager et al., 2006), involves generation of factoid questions by replacing some parts of a given proposition with abstract (i.e. ungrounded) expressions. For example, to determine the correctness of (1) and (2), two question sentences are generated for each proposition, with abstracted parts in italics and …

Content truncated to 500 words