<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-1209">
  <Title>Statistical QA - Classifier vs. Re-ranker: What's the difference?</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Open-Domain factoid Question Answering (QA) is defined as the task of answering fact-based questions phrased in Natural Language. Examples of some question and answers that fall in the fact-based category are:  1. What is the capital of Japan? - Tokyo 2. What is acetaminophen? - Non-aspirin pain killer 3. Where is the Eiffel Tower? - Paris  The architecture of most of QA systems consists of two basic modules: the information retrieval (IR) module and the answer pinpointing module. These two modules are used in a typical pipeline architecture.</Paragraph>
    <Paragraph position="1"> For a given question, the IR module finds a set of relevant segments. Each segment typically consists of at most R sentences1. The answer pinpointing module processes each of these segments and finds the appropriate answer phrase. 1 In our experiments we use R=1 phrase. Evaluation of a QA system is judged on the basis on the final output answer and the corresponding evidence provided by the segment. This paper focuses on the answer pinpointing module. Typical QA systems perform re-ranking of candidate answers as an important step in pinpointing. The goal is to rank the most likely answer first by using either symbolic or statistical methods. Some QA systems make use of statistical answer pinpointing (Xu et. al, 2002; Ittycheriah, 2001; Ittycheriah and Salim, 2002) by treating it as a classification problem. In this paper, we cast the pinpointing problem in a statistical framework and compare two approaches, classification and re-ranking.</Paragraph>
  </Section>
class="xml-element"></Paper>