<?xml version="1.0" standalone="yes"?>
<Paper uid="P01-1025">
  <Title>Methods for the Qualitative Evaluation of Lexical Association Measures</Title>
  <Section position="8" start_page="0" end_page="0" type="concl">
    <SectionTitle>
7 Conclusion
</SectionTitle>
    <Paragraph position="0"> We have shown that simple a6 -best approaches are not suitable for a qualitative evaluation of lexical association measures, mainly for the following reasons: the instability of precision values obtained from the first few percent of the data in the SLs; the lack of significant differences between the AMs after approx. 50% of the data in the SLs have been examined; and the lack of significant differences between the measures except for certain specific values of a6 . We have also shown that the evaluation results and the ranking of AMs differ depending on the kind of collocations to be identified, and the proportion of hapaxes in the candidate sets. Finally, our results question the widely accepted argument that the strength of log-likelihood lies in handling low-frequency data. In our experiments, none of the AMs was able to extract a substantial number of collocations from the set of hapaxlegomena.</Paragraph>
  </Section>
class="xml-element"></Paper>