<?xml version="1.0" standalone="yes"?>
<Paper uid="P98-2128">
  <Title>Learning Constraint Grammar-style disambiguation rules using Inductive Logic Programming</Title>
  <Section position="7" start_page="777" end_page="778" type="evalu">
    <SectionTitle>
5 Discussion and future work
</SectionTitle>
    <Paragraph position="0"> The figures of the experimental tagger are not optimal, but promising, considering that the  rules induced is a limited subset of possible rule types.</Paragraph>
    <Paragraph position="1"> Part of the explanation for the figure of ambiguities pending after tagging is that there are some ambiguity classes which are very hard to deal with. For example, there is a tag for the adverb, hB, and one tag for the verbal particle, PL. In the lexicon built from the corpus, there are 83 word forms which can have at least both these readings. Thus, turning a corpus into a lexicon might lead to the introduction of ambiguities hard to solve. A lexicon better tailored to the task would be of much use. Another important issue is that of handling unknown words.</Paragraph>
    <Paragraph position="2"> To reduce the error rate, the bad rules should be identified by testing all rules against the training data. To tackle the residual ambiguities, the next step will be to learn also different kinds of rules, for example 'select' rules which retain a given reading, but discard all others.</Paragraph>
    <Paragraph position="3"> Also rules scoping longer contexts than a window of 5-7 words must be considered.</Paragraph>
  </Section>
class="xml-element"></Paper>