<?xml version="1.0" standalone="yes"?>
<Paper uid="N01-1029">
  <Title>References</Title>
  <Section position="5" start_page="0" end_page="0" type="concl">
    <SectionTitle>
5 Conclusion and future work
</SectionTitle>
    <Paragraph position="0"> The PLCG-based SLM exposes a slight loss of robustness in the reduced recognition rate when it is used as a stand-alone rescoring LM. Combined with a word trigram LM however, perplexity and WER reductions with respect to a word 3-gram baseline seem similar to those obtained with the Chelba-Jelinek SLM and those previously reported by Chelba (2000). On the other hand, the PLCG-based SLM is significantly faster and obtains a higher parsing accuracy.</Paragraph>
    <Paragraph position="1"> In the future we plan to evaluate full EM reestimation of the models on the trainset using the formulas given in this paper.</Paragraph>
  </Section>
class="xml-element"></Paper>