<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-3807">
  <Title>Learning of Graph-based Question Answering Rules</Title>
  <Section position="6" start_page="42" end_page="43" type="concl">
    <SectionTitle>
6 Conclusions
</SectionTitle>
    <Paragraph position="0"> We have presented a method to learn question answering rules by applying graph manipulation methods on the representations of questions and answer sentences. The method is independent of the actual graph representation formalism.</Paragraph>
    <Paragraph position="1"> We are studying to combine WordNet with a Named Entity Recogniser to produce generalised rules. This way it becomes possible to replace vertices with vertex types (e.g. &amp;quot;PERSON&amp;quot;, &amp;quot;DATE&amp;quot;, etc). We are also exploring the use of machine learning techniques to learn classes of vertices. In particular, grammar induction techniques (van Zaanen, 2002) could be applied to learn types of regularities in the strings.</Paragraph>
    <Paragraph position="2">  Further research will also focus on developing methods to extend the question pattern Rp with information found in the question only. A possibility is to keep a database of question subgraphs that are allowed to be added to Rp. This database could be built by hand, but ideally it should be learnt automatically. null  Additionalresearcheffortswillbeallocatedtodetermine degrees of word similarity or paraphrasing, such as the connection between was born in and 's birthplace is. In particular, we will explore the use of nominalisations. We will also study paraphrasing methods to detect these connections.</Paragraph>
    <Paragraph position="3"> Considering that text information as complex as syntactic information or even logic and semantic information can be expressed in graphs (Quillian, 1968; Schank, 1972; Sowa, 1979), we are convinced that the time is ripe to explore the use of graphs for question answering.</Paragraph>
  </Section>
class="xml-element"></Paper>