<?xml version="1.0" standalone="yes"?>
<Paper uid="P05-1048">
  <Title>Word Sense Disambiguation vs. Statistical Machine Translation</Title>
  <Section position="2" start_page="0" end_page="387" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Word sense disambiguation or WSD, the task of determining the correct sense of a word in context, is a much studied problem area with a long and honorable history. Recent years have seen steady accuracy gains in WSD models, driven in particular by controlled evaluations such as the Senseval series of workshops. Word sense disambiguation is often assumed to be an intermediate task, which should then help higher level applications such as machine 1The authors would like to thank the Hong Kong Research Grants Council (RGC) for supporting this research in part through grants RGC6083/99E, RGC6256/00E, and DAG03/04.EG09, and several anonymous reviewers for insights and suggestions.</Paragraph>
    <Paragraph position="1"> translation or information retrieval. However, WSD is usually performed and evaluated as a standalone task, and to date there have been very few efforts to integrate the learned WSD models into full statistical MT systems.</Paragraph>
    <Paragraph position="2"> An energetically debated question at conferences over the past year is whether even the new state-of-the-art word sense disambiguation models actually have anything to offer to full statistical machine translation systems. Among WSD circles, this can sometimes elicit responses that border on implying that even asking the question is heretical. In efforts such as Senseval we tend to regard the construction of WSD models as an obviously correct, if necessarily simplified, approach that will eventually lead to essential disambiguation components within larger applications like machine translation.</Paragraph>
    <Paragraph position="3"> There is no question that the word sense disambiguation perspective has led to numerous insights in machine translation, even of the statistical variety. It is often simply an unstated assumption that any full translation system, to achieve full performance, will sooner or later have to incorporate individual WSD components.</Paragraph>
    <Paragraph position="4"> However, in some translation architectures and particularly in statistical machine translation (SMT), the translation engine already implicitly factors in many contextual features into lexical choice. From this standpoint, SMT models can be seen as WSD models in their own right, albeit with several major caveats.</Paragraph>
    <Paragraph position="5"> But typical statistical machine translation models only rely on a local context to choose among lexical translation candidates, as discussed in greater detail later. It is therefore often assumed that dedicated WSD-based lexical choice models, which can incor- null porate a wider variety of context features, can make better predictions than the &amp;quot;weaker&amp;quot; models implicit in statistical MT, and that these predictions will help the translation quality.</Paragraph>
    <Paragraph position="6"> Nevertheless, this assumption has not been empirically verified, and we should not simply assume that WSD models can contribute more than what the SMT models perform. It may behoove us to take note of the sobering fact that, perhaps analogously, WSD has yet to be conclusively shown to help information retrieval systems after many years of attempts. null In this work, we propose to directly investigate whether word sense disambiguation--at least as it is typically currently formulated--is useful for statistical machine translation. We tackle a real Chinese to English translation task using a state-of-the-art supervised WSD system and a typical SMT model. We show that the unsupervised SMT model, trained on parallel data without any manual sense annotation, yields higher BLEU scores than the case where the SMT model makes use of the lexical choice predictions from the supervised WSD model, which are more expensive to create. The reasons for the surprising difficulty of improving over the translation quality of the SMT model are then discussed and analyzed.</Paragraph>
  </Section>
class="xml-element"></Paper>