<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-0313">
  <Title>Translation Spotting for Translation Memories</Title>
  <Section position="7" start_page="0" end_page="0" type="concl">
    <SectionTitle>
5 Conclusion
</SectionTitle>
    <Paragraph position="0"> We have presented different translation spottings methods, specifically adapted to a sub-sentential translation memory system that proposes TL translations for SL sequences of syntactic chunks, as proposed by Planas (2000). These methods are based on IBM statistical translation Model 2 (Brown et al., 1993), but take advantage of certain characteristics of the segments of text that can typically be extracted from translation memories. By imposing contiguity and compositionality constraints on the search procedure, we have shown that it is possible to perform translation spotting more accurately than by simply relying on the most likely word alignment.</Paragraph>
    <Paragraph position="1"> Yet, the accuracy of our methods still leave a lot to be desired; on closer examination most of our problems can be attributed to the underlying translation model. Computing word alignments with IBM Model 2 is straightforward and efficient, which made it a good choice for experimenting; however, this model is certainly not the state of the art in statistical translation modeling. Thenagain, the methods proposed here were all based on the idea of finding the most likely word-alignment under various constraints. This approach is not dependent on the underlying translation model, and similar methods could certainly be devised based on more elaborate models, such as IBM Models 3-5, or the HMM-based models proposed by Och et al. (1999) for example.</Paragraph>
    <Paragraph position="2"> Alternatively, there are other ways to compensate for Model 2's weaknesses. Each IBM-style alignment between two segments of text denotes one particular explanation of how the TL words emerged from the SL words, but it doesn't tell the whole story. Basing our TS methods on a set of likely alignments rather than on the single most-likely alignment, as is normally done to estimate the parameters of higher-level models, could possibly lead to more accurate TS results. Similarly, TS applications are not bound to translation directionality as statistical translation systems are; this means that we could also make use of a &amp;quot;reverse&amp;quot; model to obtain a better estimate of the likelihood of two segments of text being mutual translation. null These are all research directions that we are currently pursuing.</Paragraph>
  </Section>
class="xml-element"></Paper>