Learning of Graph Rules for Question Answering

Proceedings of the Australasian Language Technology Workshop 2005, pages 15–23, Sydney, Australia, December 2005. Learning of Graph Rules for Question Answering Diego MOLLA and Menno VAN ZAANEN Centre for Language Technology, Macquarie University Sydney, Australia, {diego,menno}@ics.mq.edu.au Abstract AnswerFinder is a framework for the devel- opment of question-answering systems. An- swerFinder is currently being used to test the applicability of graph representations for the de- tection and extraction of answers. In this paper we briefly describe AnswerFinder and introduce our method to learn graph patterns that link questions with their corresponding answers in arbitrary sentences. The method is based on the translation of the logical forms of questions and answer sentences into graphs, and the ap- plication of operations based on graph overlaps and the construction of paths within graphs. The method is general and can be applied to any graph-based representation of the contents of questions and answers. 1 Introduction Text-based question answering (henceforth QA) is the process whereby an answer to an arbi- trary question formulated in plain English is found by searching through unedited text doc- uments and returned to the user. The current availability of increasingly large volumes of text for human consumption has prompted an inten- sive research in QA. A well-known forum for the evaluation of QA systems is the question- answering track of the Text REtrieval Confer- ence1 (Voorhees, 2001), where systems devel- oped by some of the most active researchers in the area are compared within the context of a common task. In addition, QA technology is being deployed in practical applications. For example, several Web-based question-answering systems are currently available (e.g. START2, AnswerBus3), and recently popular Web search engines have started incorporating automated 1http://trec.nist.gov 2http://www.ai.mit.edu/projects/infolab/ 3http://www.answerbus.com/index.shtml question-answering techniques (e.g. Google4, as for September 2005). The development of successful QA technol- ogy requires solid foundations both in the ar- eas of software engineering and natural lan- guage processing. The nature of text-based question answering requires the use of a wide range of techniques, some of which are described in (Hirschman and Gaizauskas, 2001; Voorhees, 2001). For example, traditional document re- trieval techniques are typically used to prese- lect the documents or document fragments that may contain the answer to the question. In addition, information extraction techniques are commonly used to extract all the named enti- ties in the question and the preselected text, on the ground that fact-based questions typi- cally expect one of these named entities as an answer. To analyse the questions, techniques range from the use of regular expressions to the use of machine-learning techniques that classify the questions according to the type of the ex- pected answer. Finally, to find the answer, tech- niques may vary from a bag of words compar- ison of keywords used in the question and the answer sentence, to the use of full parsers and logical proof tools such as OTTER5. Additional resources are typically used, notably the Word- Net lexical resource.6 Consequently, the most successful QA systems are complex pieces of en- gineering …

Content truncated to 500 words