<?xml version="1.0" standalone="yes"?>
<Paper uid="C04-1036">
  <Title>Feature Vector Quality and Distributional Similarity</Title>
  <Section position="9" start_page="0" end_page="79" type="ackno">
    <SectionTitle>
2. A thorough empirical error analysis of state of
</SectionTitle>
    <Paragraph position="0"> the art performance was conducted. The main observation was deficient quality of the feature vectors which reduces the quality of similarity measures.</Paragraph>
    <Paragraph position="1"> 3. Inspired by the qualitative observations we identified a new qualitative condition for feature vector evaluation - top joint feature rank. Thus, feature vector quality can be measured independently of the final similarity output.</Paragraph>
    <Paragraph position="2"> 4. Finally, we presented a novel feature weighting function, relative feature focus. This measure was designed based on error analysis insights and im- null economy along with their corresponding ranks in the two (sorted) feature vectors.</Paragraph>
    <Paragraph position="3"> proves performance over all the above criteria. We intend to further investigate the contribution of our measure to word sense disambiguation and to evaluate its performance for clustering methods. Error analysis suggests that it might be difficult to improve similarity output further within the common distributional similarity schemes. We need to seek additional criteria and data types, such as identifying evidence for non-similarity, or analyzing more carefully disjoint features.</Paragraph>
    <Paragraph position="4"> Further research is suggested to extend the learning framework towards richer notions of ontology generation. We would like to distinguish between different ontological relationships that correspond to the substitutability criterion, such as identifying the entailment direction, which was ignored till now. Towards these goals we plan to investigate combining unsupervised distributional similarity with supervised methods for learning ontological relationships, and with paraphrase acquisition methods.</Paragraph>
  </Section>
class="xml-element"></Paper>