<?xml version="1.0" standalone="yes"?>
<Paper uid="W99-0909">
  <Title>Unsupervised Lexical Learning with Categorial Grammars</Title>
  <Section position="7" start_page="361" end_page="361" type="concl">
    <SectionTitle>
6 Conclusions
</SectionTitle>
    <Paragraph position="0"> We have presented an unsupervised learner that is able to both learn CG lexicons and annotate natural language corpora, with less background knowledge than other systems in the literature.</Paragraph>
    <Paragraph position="1"> Results from preliminary experiments are encouraging with respect to both problems, particularly as the system appears to be reasonably effective on small, sparse corpora. It is encouraging that where errors arose this was often due only to incomplete background knowledge.</Paragraph>
    <Paragraph position="2"> The results presented are encouraging with respect to the work that has already been mentioned - 100~ can clearly not be improved upon and compares very favourable with the systems mentioned in Section 1. However, it is also clear that this was achieved on unrealistically simple corpora and when the system was used on the more diverse LLL corpus it did not fair as well. However, given the fact that the problem setting discussed here is somewhat harder than that attempted by other systems and the lack of linguistic background knowledge supplied, it is hoped that it will be possible to use the approach on wider coverage corpora more effectively in the future.</Paragraph>
    <Paragraph position="3"> The use of CGs to solve the problem provides an elegant way of using syntactic information to constrain the learning problem and provides the opportunity for expansion to a full grammar learning system in the future by the development of a category hypothesizer. It is hoped that this will be part of future work.</Paragraph>
    <Paragraph position="4"> .We also hope to carry out experiments on larger and more diverse corpora, as the corpora used thus far are too small to be a an exacting test for the approach. We need to expand the grammar to cover more linguistic phenomena to achieve this, as well as considering other measures for compressing the lexicon (e.g. using an MDL-based approach). Larger experiments will lead to a need for increased efficiency in the parsing and reparsing processes. This could be done by considering deterministic parsing approaches (Marcus, 1980), or perhaps shallower syntactic analysis.</Paragraph>
    <Paragraph position="5"> While many extensions may be considered for this work, the evidence thus far suggests that the approach outlined in this paper is effective and efficient for these natural language learning tasks.</Paragraph>
  </Section>
class="xml-element"></Paper>