<?xml version="1.0" standalone="yes"?>
<Paper uid="W02-1906">
  <Title>Passage Selection to Improve Question Answering</Title>
  <Section position="6" start_page="1" end_page="1" type="concl">
    <SectionTitle>
5 Conclusions and future works
</SectionTitle>
    <Paragraph position="0"> In this paper, we have analysed the improvement obtained by our passage retrieval system, called IR-n, with reference to a high-performance IR system (ATT) regarding to is application for QA tasks. This improvement has been evaluated on the TREC-9 QA track data set. The achieved improvements are twofold: First, our approach obtains a better precision by retrieving more passages that contain the answer to users' queries than ATT system does. Second, since our approach returns passages (instead of documents), it significantly reduces the amount of text to be processed with costly techniques by the QA system. The related experiments show that the optimal passage length for this task is 20 when passages are made up by a fixed number of sentences. Moreover, we have tested two different ways of applying our model. As we have seen, IR-n presents similar results when it works on the output of an IR system, than when it works on the whole collections. Nevertheless, in both cases, benefits range from 6 to 12 points with reference to ATT system depending on the number of first documents or passages retrieved to be processed for QA tasks.</Paragraph>
    <Paragraph position="1"> As future work, in order to improve our system precision, we intend to obtain the optimum size of passages in accordance with the kind of question. Besides, we need to investigate the effects of query expansion techniques on IR-n system. Furthermore, we are also trying to improve the relationship between IR-n and the following QA system, in order to detect the minimum number of passages to extract for each query without affecting QA performance.</Paragraph>
  </Section>
class="xml-element"></Paper>