<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-1203">
  <Title>Automatic Identification of Non-Compositional Multi-Word Expressions using Latent Semantic Analysis</Title>
  <Section position="6" start_page="16" end_page="16" type="concl">
    <SectionTitle>
4 Conclusion
</SectionTitle>
    <Paragraph position="0"> To summarize, in order to classify an MWE as non-compositional, we compute an approximation of its compositional meaning and compare this with the meaning of the expression as it is used on the whole. One of the obvious improvements to the algorithm could come from better models for simulating compositional meaning. A further issue that can be explored is whether linguistic preprocessing would influence the results. We worked only on raw text data. There is some evidence (Baldwin et al., 2003) that part of speech tagging might improve results in this kind of task.</Paragraph>
    <Paragraph position="1"> We also only considered local word sequences.</Paragraph>
    <Paragraph position="2"> Certainly some recognition of the syntactic structure would improve results. These are, however, more general issues associated with MWE processing. null Rather promising results were attained using only local context, however. Our study shows that the F-score measure is maximized by taking as threshold for distinguishing non-compositional phrases from compositional ones a cosine similarity value somewhere between 0.1-0.2. An important point to be explored is that compositionality appears to come in degrees. As Bannard and Lascarides (2003) have noted, MWEs &amp;quot;do not fall cleanly into the binary classes of compositional and non-compositional expressions, but populate a continuum between the two extremes.&amp;quot; While our experiment was designed to classify MWEs, the technique described here, of course, provides a means, if rather a blunt one, for quantifying the degreee of compositonality of an expression.</Paragraph>
  </Section>
class="xml-element"></Paper>