<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-0831">
  <Title>Novel Reordering Approaches in Phrase-Based Statistical Machine Translation</Title>
  <Section position="4" start_page="0" end_page="167" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Reordering is of crucial importance for machine translation. Already (Knight et al., 1998) use full unweighted permutations on the level of source words in their early weighted finite-state transducer approach which implemented single-word based translation using conditional probabilities. In a refinement with additional phrase-based models, (Kumar et al., 2003) define a probability distribution over all possible permutations of source sentence phrases and prune the resulting automaton to reduce complexity. null A second category of finite-state translation approaches uses joint instead of conditional probabilities. Many joint probability approaches originate in speech-to-speech translation as they are the natural choice in combination with speech recognition models. The automated transducer inference techniques OMEGA (Vilar, 2000) and GIATI (Casacuberta et al., 2004) work on phrase level, but ignore the re-ordering problem from the view of the model. Without reordering both in training and during search, sentences can only be translated properly into a language with similar word order. In (Bangalore et al., 2000) weighted reordering has been applied to target sentences since defining a permutation model on the source side is impractical in combination with speech recognition. In order to reduce the computational complexity, this approach considers only a set of plausible reorderings seen on training data.</Paragraph>
    <Paragraph position="1"> Most other phrase-based statistical approaches like the Alignment Template system of Bender et al. (2004) rely on (local) reorderings which are implicitly memorized with each pair of source and target phrases in training. Additional reorderings on phrase level are fully integrated into the decoding process, which increases the complexity of the system and makes it hard to modify. Zens et al. (2003) reviewed two types of reordering constraints for this type of translation systems.</Paragraph>
    <Paragraph position="2"> In our work we follow a phrase-based translation approach, applying source sentence reordering on word level. We compute a reordering graph on-demand and take it as input for monotonic translation. This approach is modular and allows easy introduction of different reordering constraints and probabilistic dependencies. We will show that it performs at least as well as the best statistical machine translation system at the IWSLT Evaluation.</Paragraph>
    <Paragraph position="3">  In the next section we briefly review the basic theory of our translation system based on weighted finite-state transducers (WFST). In Sec. 3 we introduce new methods for reordering and alignment monotonization in training. To compare different reordering constraints used in the translation search process we develop an on-demand computable framework for permutation models in Sec. 4.</Paragraph>
    <Paragraph position="4"> In the same section we also define and analyze unrestricted and restricted permutations with some of them being first published in this paper. We conclude the paper by presenting and discussing a rich set of experimental results.</Paragraph>
  </Section>
class="xml-element"></Paper>