<?xml version="1.0" standalone="yes"?>
<Paper uid="W02-1032">
  <Title>Exploiting Headword Dependency and Predictive Clustering for Language Modeling</Title>
  <Section position="9" start_page="7" end_page="7" type="concl">
    <SectionTitle>
7 Conclusion
</SectionTitle>
    <Paragraph position="0"> We proposed and evaluated a new language model, the permuted headword trigram model with clustering (C-PHTM). We have shown that the simple model that combines the predictive clustering with a headword detector can effectively capture structure in language. Experiments show that the proposed model achieves an encouraging 15% CER reduction over a conventional word trigram model in a Japanese Kana-Kanji conversion system. We also compared C-PTHM to several similar models, showing that our model has many practical advantages, and achieves substantially better performance.</Paragraph>
    <Paragraph position="1"> One issue we did not address in this paper was the language model size: the models that use HTM are larger than the baseline model we compared the performance with. Though we did not pursue the issue of size reduction in this paper, there are many known techniques that effectively reduce the model size while minimizing the loss in performance. One area of future work is therefore to reduce the model size. Other areas include the application of the proposed model to a wider variety of test corpora and to related tasks.</Paragraph>
  </Section>
class="xml-element"></Paper>