<?xml version="1.0" standalone="yes"?>
<Paper uid="E06-1052">
  <Title>Investigating a Generic Paraphrase-based Approach for Relation Extraction</Title>
  <Section position="8" start_page="414" end_page="415" type="concl">
    <SectionTitle>
7 Conclusions and Future Work
</SectionTitle>
    <Paragraph position="0"> We have presented a paraphrase-based approach for relation extraction (RE), and an implemented system, that rely solely on unsupervised paraphrase acquisition and generic syntactic template matching. Two targets were investigated: (a) a mostly unsupervised, domain independent, configuration for RE, and (b) an evaluation scheme for paraphrase acquisition, providing a first evaluationofitsrealisticcoverage. Ourapproachdiffers from previous unsupervised IE methods in that we identify instances of a specific relation while prior methods identified template relevance only at the general scenario level.</Paragraph>
    <Paragraph position="1"> We manually analyzed the potential of our approach on a dataset annotated with protein interactions. The analysis shows that 93% of the interacting protein pairs can be potentially identified with the template-based approach. Addi- null tionally, we manually assessed the coverage of the TEASE acquisition algorithm and found that 63% of the distinct pairs can be potentially recognized with the learned templates, assuming an ideal matcher, indicating a significant potential recall for completely unsupervised paraphrase acquisition. Finally, weevaluated our currentsystem performance and found it weaker than supervised RE methods, being far from fulfilling the potential indicated in our manual analyses due to insufficient syntactic matching. But, even our current performance may be considered useful given the very small amount of domain-specific information used by the system.</Paragraph>
    <Paragraph position="2"> Most importantly, we believe that our analysis and evaluation methodologies for an RE dataset provide an excellent benchmark for unsupervised learning of paraphrases and entailment rules. In the long run, we plan to develop and improve our acquisition and matching algorithms, in order to realize the observed potential of the paraphrase-based approach. Notably, our findings point to the need to learn generic morphological and syntactic variations in template matching, an area which has rarely been addressed till now.</Paragraph>
  </Section>
class="xml-element"></Paper>