<?xml version="1.0" standalone="yes"?>
<Paper uid="W99-0607">
  <Title>Applying Extrasentential Context To Maximum Entropy Based Tagging With A Large Semantic And Syntactic Tagset</Title>
  <Section position="10" start_page="82425" end_page="82425" type="concl">
    <SectionTitle>
6 Conclusion
</SectionTitle>
    <Paragraph position="0"> Our main concern in this paper has been to show that extrasentential information can provide significant assistance to a real tagger. There has been almost no research done in this area, possibly due to the fact that, for small syntax-only tagsets, very accurate performance can be obtained labelling the Wall Street Journal corpus using only local context. In the experiments presented, we have used a much more detailed, semantic and syntactic tagset, on which the performance is much lower. Extrasentential semantic information is needed to disambiguate these tags. We have observed that the simple approach of only using the occurrence of tags in the history as features did not significantly improve performance. However, when more sophisticated questions are employed to mine this long-range contextual information, a more significant contribution to performance is made. This motivates further research toward finding more predictive features. Clearly, the work here has only scratched the surface in terms of the kinds of questions that it is possible to ask of the history. The maximum entropy approach that we have adopted is extremely accommodating in this respect. It is possible to  go much further in the direction of querying the historical tag structure. For example, we can, in effect, exploit grammatical relations within previous sentences with an eye to predicting the tags of similarly related words in the current sentence. It is also possible to go even further and exploit the structure of full parses in the history.</Paragraph>
  </Section>
class="xml-element"></Paper>