<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-3113">
  <Title>How Many Bits Are Needed To Store Probabilities for Phrase-Based Translation?</Title>
  <Section position="7" start_page="98" end_page="98" type="concl">
    <SectionTitle>
6 Conclusion
</SectionTitle>
    <Paragraph position="0"> In this paper we investigated the application of data compression methods to the probabilities stored by a phrase-based translation model. In particular, probability quantization was applied on the n-gram language model and on the phrase-pair translation model. Experimental results con rm previous ndings in speech recognition: language model probabilities can be encoded in just 4 bits at the cost of a very little loss in performance. The same resolution level seems to be a good compromise even for the translation model. Remarkably, the impact of quantization on the language model and translation model seems to be additive with respect to performance. Finally, quantization does not seems to be affected by data sparseness and behaves similarly on different translation tasks.</Paragraph>
  </Section>
class="xml-element"></Paper>