<?xml version="1.0" standalone="yes"?>
<Paper uid="H05-1115">
  <Title>Using Random Walks for Question-focused Sentence Retrieval</Title>
  <Section position="3" start_page="915" end_page="915" type="intro">
    <SectionTitle>
2% Summary
</SectionTitle>
    <Paragraph position="0"> The North Carolina coast braced for a weakened but still potent Hurricane Isabel while already rain-soaked areas as far away as Pennsylvania prepared for possibly ruinous flooding. (2:3) A hurricane warning was in effect from Cape Fear in southern North Carolina to the Virginia-Maryland line, and tropical storm warnings extended from South Carolina to New Jersey. (2:14) While the outer edge of the hurricane approached the North Carolina coast Wednesday, the center of the storm was still 400 miles south-southeast of Cape Hatteras, N.C., late Wednesday morning. (3:10) BBC NEWS World Americas Hurricane Isabel prompts US shutdown (4:1) Ask us: What states have been affected by the hurricane so far? Around 200,000 people in coastal areas of North Carolina and Virginia were ordered to evacuate or risk getting trapped by flooding from storm surges up to 11 feet. (5:8) The storm was expected to hit with its full fury today, slamming into the North Carolina coast with 105-mph winds and 45-foot wave crests, before moving through Virginia and bashing the capital with gusts of about 60 mph. (7:6)  from the classic task of PR for a Q&amp;A system in interesting ways, due to the time-sensitive nature of the stories inour corpus. Forexample, one challenge is that the answer to a user's question may be updated and reworded over time by journalists in order to keep a running story fresh, or because the facts themselves change. Therefore, there is often more than one correct answer to a question.</Paragraph>
    <Paragraph position="1"> We aim to develop a method for sentence retrieval that goes beyond finding sentences that are similar to a single query. To this end, we propose to use a stochastic, graph-based method. Recently, graph-based methods have proved useful for a number of NLP and IR tasks such as document re-ranking in ad hoc IR (Kurland and Lee, 2005) and analyzing sentiments in text (Pang and Lee, 2004). In (Erkan and Radev, 2004), we introduced the LexRank method and successfully applied it to generic, multi-document summarization. Presently, we introduce topic-sensitive LexRank in creating a sentence retrieval system. We evaluate its performance against a competitive baseline, which considers the similarity between each sentence and the question (using IDF-weighed word overlap). We demonstrate that LexRank significantly improves question-focused sentence selection over the baseline. null</Paragraph>
  </Section>
class="xml-element"></Paper>