<?xml version="1.0" standalone="yes"?>
<Paper uid="P05-1014">
  <Title>The Distributional Inclusion Hypotheses and Lexical Entailment</Title>
  <Section position="12" start_page="112" end_page="113" type="concl">
    <SectionTitle>
7 Conclusions and Future Work
</SectionTitle>
    <Paragraph position="0"> The main contributions of this paper were: 1. We defined two Distributional Inclusion Hypotheses that associate feature inclusion with lexical entailment at the word sense level. The Hypotheses were proposed as a refinement for Harris' Distributional hypothesis and as an extension to the classic distributional similarity scheme. 2. To estimate the empirical validity of the defined hypotheses we developed an automatic inclusion testing algorithm (ITA). The core of the algorithm is a web-based feature inclusion testing procedure, which helped significantly to compensate for data sparseness.</Paragraph>
    <Paragraph position="1"> 3. Then a thorough analysis of the data behavior with respect to the proposed hypotheses was conducted. The first hypothesis was almost fully attested by the data, particularly at the sense level, while the second hypothesis did not fully hold.</Paragraph>
    <Paragraph position="2"> 4. Motivated by the empirical analysis we proposed to employ ITA for the practical task of improving lexical entailment acquisition. The algorithm was applied as a filtering technique on the distributional similarity (RFF) output. We ob- null filter, with 20 and 40 feature sampling, compared to RFF top-40 and RFF top-26 similarities. ITA-20 and ITA-40 denote the web-sampling method with 20 and random 40 features, respectively.</Paragraph>
    <Paragraph position="3">  tained 17% increase of precision and succeeded to improve relative F1 by 15% over the baseline.</Paragraph>
    <Paragraph position="4"> Although the results were encouraging our manual data analysis shows that we still have to handle word ambiguity. In particular, this is important in order to be able to learn the direction of entailment. To achieve better precision we need to increase feature discriminativeness. To this end syntactic features may be extended to contain more than one word, and ways for automatic extraction of features from the web (rather than from a corpus) may be developed. Finally, further investigation of combining the distributional and the co-occurrence pattern-based approaches over the web is desired.</Paragraph>
  </Section>
class="xml-element"></Paper>