<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-3807">
  <Title>Learning of Graph-based Question Answering Rules</Title>
  <Section position="2" start_page="0" end_page="37" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Text-based question answering (QA) is the process of automatically finding the answers to arbitrary questions in plain English by searching collections of text files. Recently there has been intensive research in this area, fostered by evaluation-based conferences such as the Text REtrieval Conference (TREC) (Voorhees, 2001b), the Cross-Lingual Evaluation Forum (CLEF) (Vallin et al., 2005), and the NII-NACSIS Test Collection for Information Retrieval Systems workshops (NTCIR) (Kando, 2005).</Paragraph>
    <Paragraph position="1"> Current research focuses on factoid question answering, whereby the answer is a short string that indicates a fact, usually a named entity. An example of a factoid question is Who won the 400m race in the 2000 Summer Olympic games?, which has a short answer: Cathy Freeman.</Paragraph>
    <Paragraph position="2"> There are various approaches to question answering. The focus of this paper is on rule-based systems. A rule could be, say, &amp;quot;if the question is of the form Who is the &lt;position&gt; of &lt;country&gt;&amp;quot; and a text sentence says &lt;position&gt; of &lt;country&gt; Y and Y consists of two capitalised words, then Y is the answer&amp;quot;). Such a rule was used by Soubbotin (2001), who developed a system who obtained the best accuracy in the 2001 Text REtrieval Conference (Voorhees, 2001a). The system developed by Soubbotin (2001) relied on the development of a large set of patterns of potential answer expressions, and the allocation of those patterns to types of questions. The patterns were developed by hand by examining the data.</Paragraph>
    <Paragraph position="3"> Soubbotin (2001)'s work shows that a rule-based QA system can produce good results if the rule set is comprehensive enough. Unfortunately, if the system is ported to a new domain the set of rules needs to be ported as well. It has not been proven that rules like the ones developed by Soubbotin (2001), which were designed for the TREC QA task, can be ported to other domains. Furthermore, the process of producing the rules was presumably very labour intensive. Consequently, the cost of manually producing newrulesforaspecialiseddomaincouldbecometoo expensive for some domains.</Paragraph>
    <Paragraph position="4"> In this paper we present a method for the automatic learning of question answering rules by applying graph manipulation methods. The method relies on the representation of questions and answer sentences as graphs. Section 2 describes the general format of the graph-based QA rules and section 3 describes the method to learn the rules. The methodsdescribedontheabovetwosectionsareindepen- null dentoftheactualsentencerepresentationformalism,  as long as the representation is a graph. Section 4 presents a specific application using logical graphs.</Paragraph>
    <Paragraph position="5"> Finally, sections 5 and 6 focus on related research and final conclusions, respectively.</Paragraph>
  </Section>
class="xml-element"></Paper>