<?xml version="1.0" standalone="yes"?>
<Paper uid="W98-1304">
  <Title>I</Title>
  <Section position="11" start_page="75" end_page="75" type="concl">
    <SectionTitle>
10 Discussion
</SectionTitle>
    <Paragraph position="0"> The method we propose analyses language indirectly as a regular language. This makes it impossible to use long distance dependencies, but nevertheless the experiment shows that it  performs quite reasonable and is very robust.</Paragraph>
    <Paragraph position="1"> The score is less than the scores obtained by systems that consider the entire sentence with, in particular, the headwords of phrases. But the method creates new possibilities such as processing ungrammatical text and processing unpunctuated text. Shallow parsing is also a possible application.</Paragraph>
    <Paragraph position="2"> As far as future directions are concerned, we would like to mention that our parsing strategy is not limited, to regular languages and HMM models. It is possible to switch to a history-based approach, where the choice of si depends on both the words wl...w~ and all earlier tags and traversal strings chosen by the system. In that case a statistical decision tree or a markov field can be used to model the optimal choice for s~ after seeing word wi.</Paragraph>
  </Section>
class="xml-element"></Paper>