<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-3126">
  <Title>The LDV-COMBO system for SMT</Title>
  <Section position="5" start_page="167" end_page="168" type="concl">
    <SectionTitle>
4 Conclusions and Further Work
</SectionTitle>
    <Paragraph position="0"> Many researchers remain sceptical about the usefulness of linguistic information in SMT, because, except in a couple of cases (Charniak et al., 2003; Collins et al., 2005), little success has been reported.</Paragraph>
    <Paragraph position="1"> In this work we have shown that liniguistic information may be helpful, specially when the target language has a rich morphology (e.g. Spanish).</Paragraph>
    <Paragraph position="2"> Moreover, it has often been argued that linguistic information does not yield significant improvements in MT quality, because (i) linguistic processors introduce many errors and (ii) the BLEU score is not specially sensitive to the grammaticality of MT output. We have minimized the impact of the first argument by using highly accurate tools for both languages. In order to solve the second problem more sophisticated metrics are required. Current MT evaluation metrics fail to capture many aspects of MT  quality that characterize human translations with respect to those produced by MT systems. We are devoting most of our efforts to the deployment of a new MT evaluation framework which allows to combine several similarity metrics into a single measure of quality (Gim'enez and Amig'o, 2006).</Paragraph>
    <Paragraph position="3"> We also leave for further work the experimentation of new data views such as word senses and semantic roles, as well as their natural porting from the alignment step to phrase extraction and decoding.</Paragraph>
  </Section>
class="xml-element"></Paper>