<?xml version="1.0" standalone="yes"?>
<Paper uid="N04-4006">
  <Title>Language model adaptation with MAP estimation and the perceptron algorithm</Title>
  <Section position="6" start_page="0" end_page="0" type="concl">
    <SectionTitle>
4 Discussion
</SectionTitle>
    <Paragraph position="0"> This paper has presented a series of experimental results that compare using MAP estimation for language model domain adaptation to a discriminative modeling approach for correcting errors produced by an out-of-domain model when applied to the novel domain. Because the MAP estimation produces a model that is used during first pass search, it has an advantage over the perceptron algorithm, which simply re-weights paths already in the word lattice. In support of this argument, we showed that, by using a subset of the in-domain adaptation data for MAP estimation, and the rest for use in the perceptron algorithm, we achieved results at nearly the same level as MAP estimation on the entire adaptation set.</Paragraph>
    <Paragraph position="1">  ing systems obtained by supervised LM adaptation on the 17 hour adaptation set using the second method of combination of the two methods, versus the baseline out-of-domain system.</Paragraph>
    <Paragraph position="2"> With a more complicated training scenario, which used all of the in-domain adaptation data for both methods jointly, we were able to improve WER over MAP estimation alone by 0.7 percent, for a total improvement over the baseline of 8.4 percent.</Paragraph>
    <Paragraph position="3"> Studying the various options for incorporating the perceptron algorithm within the multi-pass rescoring framework, our results show that there is a benefit from incorporating the perceptron at an early search pass, as it produces more accurate transcripts for unsupervised adaptation. Furthermore, it is important to closely match testing conditions for perceptron training.</Paragraph>
  </Section>
class="xml-element"></Paper>