<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-3115">
  <Title>NTT System Description for the WMT2006 Shared Task</Title>
  <Section position="5" start_page="123" end_page="124" type="evalu">
    <SectionTitle>
4 Results
</SectionTitle>
    <Paragraph position="0"> Table 3 shows the open test translation results on 2005 and 2006 test set (the development-test set and the final test set) 2. We used the merged (hierarchical) phrase tables for decoding. Feature function scaling factors were optimized on BLEU score using the supplied development set that is identical to the 2005's development set. We observed that our 2We did not differetiated in-domain or out-of-domain for 2006 test set.</Paragraph>
    <Paragraph position="1">  results are very comparable to the last year's best results in test2005. Also found that our hierarchical phrase-based translation (Rule) performed slightly inferior to the phrase-based translation (Phrase) in both test sets. The hierarchically combined phrases seem to be too flexible to represent the relationship of similar language pairs. Note that our hierarchical phrase-based model performed better in the English-to-German translation task. Those language pair requires rather distorted reordering, which could be represented by hierarchically combined phrases.</Paragraph>
    <Paragraph position="2"> We also conducted additional studies on how differently aligned corpora might affect the translation quality on Spanish-to-English task for the 2005 test set. Using our phrase-based model, the BLEU scores for lower/stem/prefix4 were 30.90/30.89/30.76, respectively. The differences of translation qualities were statistically significant at the 95% confidence level. Our phrase translation pairs aggregated from all the differently preprocessed corpora improved the translation quality.</Paragraph>
  </Section>
class="xml-element"></Paper>