<?xml version="1.0" standalone="yes"?>
<Paper uid="H92-1021">
  <Title>IMPROVEMENTS IN STOCHASTIC LANGUAGE MODELING</Title>
  <Section position="8" start_page="108" end_page="108" type="concl">
    <SectionTitle>
4. SUMMARY AND CONCLUSIONS
</SectionTitle>
    <Paragraph position="0"> We presented two attempts to improve our stochastic language modeling. In the first, we identified a deficiency in the conventional backoff language model, and used statistical reasoning to correct it. Our modified model is about as simple as the original one, but gives a slightly lower perplexity on various tasks. Our analysis suggests that the modification is most suitable when training data is sparse.</Paragraph>
    <Paragraph position="1"> In our second attempt, we extended the notion of adaptation to incorporate within-document word sequence correlation, using the framework of a trigger pair. We discussed the issues involvedin constructing such a model, and reported promising improvements in perplexity. We have only begun to explore the potential of trigger-based adaptive models. The results reported here are preliminary. We believe we can improve our performance by implementing many of the ideas suggested in sections 3.2, 3.3 and 3.4 above. Work is already under way.</Paragraph>
  </Section>
class="xml-element"></Paper>