<?xml version="1.0" standalone="yes"?>
<Paper uid="P04-1058">
  <Title>Alternative Approaches for Generating Bodies of Grammar Rules</Title>
  <Section position="8" start_page="30" end_page="30" type="concl">
    <SectionTitle>
7 Conclusions and Future Work
</SectionTitle>
    <Paragraph position="0"> Our experiments support two kinds of conclusions.</Paragraph>
    <Paragraph position="1"> First, modeling rules with algorithms other than n-grams not only produces smaller grammars but also better performing ones. Second, the procedure used for optimizing alpha reveals that some POS behave almost deterministically for selecting their arguments, while others do not. These findings suggests that splitting classes that behave non-deterministically into homogeneous ones could improve the quality of the inferred automata. We saw that lexicalization and head-annotation seem to attack this problem. Obvious questions for future work arise: Are these two techniques the best way to split non-homogeneous classes into homogeneous ones? Is there an optimal splitting?</Paragraph>
  </Section>
class="xml-element"></Paper>