<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-3108">
  <Title>Discriminative Reordering Models for Statistical Machine Translation</Title>
  <Section position="3" start_page="0" end_page="55" type="relat">
    <SectionTitle>
2 Related Work
</SectionTitle>
    <Paragraph position="0"> As already mentioned in Section 1, many current phrase-based statistical machine translation systems use a very simple reordering model: the costs  for phrase movements are linear in the distance. This approach is also used in the publicly available Pharaoh decoder (Koehn, 2004). The idea of predicting the orientation is adopted from (Tillmann and Zhang, 2005) and (Koehn et al., 2005). Here, we use the maximum entropy principle to combine a variety of different features.</Paragraph>
    <Paragraph position="1"> A reordering model in the framework of weighted finite state transducers is described in (Kumar and Byrne, 2005). There, the movements are defined at the phrase level, but the window for reordering is very limited. The parameters are estimated using an EM-style method.</Paragraph>
    <Paragraph position="2"> None of these methods try to generalize from the words or phrases by using word classes or part-of-speech information.</Paragraph>
    <Paragraph position="3"> The approach presented here has some resemblance to the bracketing transduction grammars (BTG) of (Wu, 1997), which have been applied to a phrase-based machine translation system in (Zens et al., 2004). The difference is that, here, we do not constrain the phrase reordering. Nevertheless the inverted/monotone concatenation of phrases in the BTG framework is similar to the left/right phrase orientation used here.</Paragraph>
  </Section>
class="xml-element"></Paper>