<?xml version="1.0" standalone="yes"?>
<Paper uid="P98-1048">
  <Title>Sylvain_Delisle @uqtr.uquebec.ca</Title>
  <Section position="6" start_page="312" end_page="312" type="concl">
    <SectionTitle>
6 Conclusion
</SectionTitle>
    <Paragraph position="0"> We have presented an experiment which demonstrates that machine learning may be used as a technique to optimise in an adaptive manner the high-level decisions that any parser must make in the presence of incomplete information about the properties of the text it analyses. The results show clearly that simple and understandable rules learned by machine learning techniques can surpass the performance of heuristics supplied by an experienced computational linguist. Moreover, these very encouraging results indicate that the representation that we chose and discuss was an adequate one for this problem. We feel that a methodology is at hand to extend and deepen this approach to language processing programs in general. The methodology consists of three main steps: I) feature engineering, 2) learning, using several different available learners, 3) evaluation, with the recommendation of using the &amp;quot;out-of-sample&amp;quot; approach to testing. Future work will focus on improvements to constructive learning; on new ways of integrating the rules acquired by different learners in the parser; and on the identification of criteria for selecting parser rules that have the best potential to benefit from the generalisation of our results.</Paragraph>
  </Section>
class="xml-element"></Paper>