<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-1065">
  <Title>Sydney, July 2006. c(c)2006 Association for Computational Linguistics Improved Discriminative Bilingual Word Alignment</Title>
  <Section position="4" start_page="0" end_page="513" type="intro">
    <SectionTitle>
2 Overall Approach
</SectionTitle>
    <Paragraph position="0"> As in our previous work (Moore, 2005), we train two models we call stage 1 and stage 2, both in the form of a weighted linear combination of feature values extracted from a pair of sentences and a proposed word alignment of them. The possible alignment having the highest overall score is selected for each sentence pair. Thus, for a sentence pair (e,f) we seek the alignment ^a such that</Paragraph>
    <Paragraph position="2"> where the fi are features and the li are weights.</Paragraph>
    <Paragraph position="3"> The models are trained on a large number of bilingual sentence pairs, a small number of which have hand-created word alignments provided to the training procedure. A set of hand alignments of a different subset of the overall training corpus is used to evaluate the models.</Paragraph>
    <Paragraph position="4"> In the stage 1 model, all the features are based on surface statistics of the training data, plus the hypothesized alignment. The entire training corpus is then automatically aligned using this model. The stage 2 model uses features based not only on the parallel sentences themselves but also on statistics of the alignments produced by the stage  1 model. The stage 1 model is discussed in Section 3, and the stage 2 model, in Section 4. After experimenting with many features and combinations of features, we made the final selection based on minimizing training set AER.</Paragraph>
    <Paragraph position="5"> For alignment search, we use a method nearly identical to our previous beam search procedure, which we do not discuss in detail. We made two minor modifications to handle the possiblity that more than one alignment may have the same score, which we previously did not take into account.</Paragraph>
    <Paragraph position="6"> First, we modified the beam search so that the beam size dynamically expands if needed to accomodate all the possible alignments that have the same score. Second we implemented a structural tie breaker, so that the same alignment will always be chosen as the one-best from a set of alignments having the same score. Neither of these changes significantly affected the alignment results.</Paragraph>
    <Paragraph position="7"> The principal training method is an adaptation of averaged perceptron learning as described by Collins (2002). The differences between our current and earlier training methods mainly address the observation that perceptron training is very sensitive to the order in which data is presented to the learner. We also investigated the large-margin training technique described by Tsochantaridis et al. (2004). The training procedures are described in Sections 5 and 6.</Paragraph>
  </Section>
class="xml-element"></Paper>