<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-1002">
  <Title>Going Beyond AER: An Extensive Analysis of Word Alignments and Their Impact on MT</Title>
  <Section position="4" start_page="9" end_page="9" type="relat">
    <SectionTitle>
2 Related Work
</SectionTitle>
    <Paragraph position="0"> Starting with the IBM models (Brown et al., 1993), researchers have developed various statistical word alignment systems based on different models, such as hidden Markov models (HMM) (Vogel et al., 1996), log-linear models (Och and Ney, 2003), and similarity-based heuristic methods (Melamed, 2000). These methods are unsupervised, i.e., the only input is large parallel corpora. In recent years, researchers have shown that even using a limited amount of manually aligned data improves word alignment significantly (Callison-Burch et al., 2004). Supervised learning techniques, such as perceptron learning, maximum entropy modeling or maximum weighted bipartite matching, have been shown to provide further improvements on word alignments (Ayan et al., 2005; Moore, 2005; Ittycheriah and Roukos, 2005; Taskar et al., 2005).</Paragraph>
    <Paragraph position="1"> The standard technique for evaluating word alignments is to represent alignments as a set of links (i.e., pairs of words) and to compare the generated alignment against manual alignment of the same data at the level of links. Manual alignments are represented by two sets: Probable (P) alignments and Sure (S) alignments, where S [?] P. Given A,P and S, the most commonly used metrics--precision (Pr), recall (Rc) and alignment error rate (AER)--are defined as follows:</Paragraph>
    <Paragraph position="3"> Another approach to evaluating alignments is to measure their impact on an external application, e.g., statistical MT. In recent years, phrase-based systems (Koehn, 2004; Chiang, 2005) have been shown to outperform word-based MT systems; therefore,inthispaper,weuseapublicly-available phrase-based MT system, Pharaoh (Koehn, 2004), to investigate the impact of different alignments.</Paragraph>
    <Paragraph position="4"> Although it is possible to estimate phrases directly from a training corpus (Marcu and Wong, 2002), most phrase-based MT systems (Koehn, 2004; Chiang, 2005) start with a word alignment and extract phrases that are consistent with the given alignment. Once the consistent phrases are extracted, they are assigned multiple scores (such  as translation probabilities and lexical weights), and the decoder's job is to choose the correct phrases based on those scores using a log-linear model.</Paragraph>
  </Section>
class="xml-element"></Paper>