<?xml version="1.0" standalone="yes"?>
<Paper uid="P05-1057">
  <Title>Log-linear Models for Word Alignment</Title>
  <Section position="11" start_page="464" end_page="465" type="concl">
    <SectionTitle>
3 E - C (MEC); l2: Model 3 C - E (MCE); l3:
</SectionTitle>
    <Paragraph position="0"> POS E - C (PEC); l4: POS C - E (PCE); l5: Dict (normalized such that summationtext5m=1 lm = 1).</Paragraph>
    <Paragraph position="1"> results show that log-linear models for word alignment significantly outperform IBM translation models. However, the search algorithm we proposed is  supervised, relying on a hand-aligned bilingual corpus, while the baseline approach of IBM alignments is unsupervised.</Paragraph>
    <Paragraph position="2"> Currently, we only employ three types of knowledge sources as feature functions. Syntax-based translation models, such as tree-to-string model (Yamada and Knight, 2001) and tree-to-tree model (Gildea, 2003), may be very suitable to be added into log-linear models.</Paragraph>
    <Paragraph position="3"> It is promising to optimize the model parameters directly with respect to AER as suggested in statistical machine translation (Och, 2003).</Paragraph>
  </Section>
class="xml-element"></Paper>