<?xml version="1.0" standalone="yes"?>
<Paper uid="H93-1021">
  <Title>ADAPTIVE LANGUAGE MODELING USING THE MAXIMUM ENTROPY PRINCIPLE</Title>
  <Section position="10" start_page="111" end_page="112" type="concl">
    <SectionTitle>
9. RESULTS
</SectionTitle>
    <Paragraph position="0"> The ML/ME model described above was trained on 5 million words of Wail Street Journal text, using DARPA's official &amp;quot;200&amp;quot; vocabulary of some 20,000 words. A conventionai trigram model was used as a baseline. The constraints used by the ML/ME model were: 18,400 unigram constraints, 240,000 bigram constraints, and 414,000 trigram constraints.</Paragraph>
    <Paragraph position="1"> One experiment was run with 36,000 trigger constraints (best 3 triggers for each word), and another with 65,000 trigger constraints (best 6 triggers per word). All models were trained on the same data, and evaluated on 325,000 words on independent data. The Maximum Entropy models were also interpolated with the conventional trigram, using yet unseen data for interpolation. Results are summarized in table 1.</Paragraph>
    <Paragraph position="2">  Entropy model over a conventional trigram model. Training is on 5 million words of WSJ text. Vocabulary is 20,000 words.</Paragraph>
    <Paragraph position="3"> The trigger constraints used in this run were selected very crudely, and their number was not optimized. We believe much more improvement can be achieved. Special modeling of self triggers has not been implemented yet. Similarly, we expect it to yield further improvement.</Paragraph>
  </Section>
class="xml-element"></Paper>