<?xml version="1.0" standalone="yes"?>
<Paper uid="P93-1022">
  <Title>CONTEXTUAL WORD SIMILARITY AND ESTIMATION FROM SPARSE DATA</Title>
  <Section position="8" start_page="169" end_page="169" type="concl">
    <SectionTitle>
6 Conclusions
</SectionTitle>
    <Paragraph position="0"> In both evaluations, similarity based estimation performs better than frequency based estimation.</Paragraph>
    <Paragraph position="1"> This indicates that when trying to estimate cooccurrence probabilities, it is useful to consider the cooccurrence patterns of the specific words and not just their frequencies, as smoothing methods do. Comparing with class based models, our approach suggests the advantage of making the most specific analogies for each word, instead of making analogies with all members of a class, via general class parameters. This raises the question whether generalizations over word classes, which follow long traditions in semantic classification, indeed provide the best means for inferencing about properties of words.</Paragraph>
  </Section>
class="xml-element"></Paper>