<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-1415">
  <Title>Generating Intelligent Numerical Answers in a Question-Answering System</Title>
  <Section position="4" start_page="0" end_page="103" type="intro">
    <SectionTitle>
1.1 Related work
</SectionTitle>
    <Paragraph position="0"> Search engines on the Web produce a set of answers to a question in the form of hyperlinks or page extracts, ranked according to content or popularity criteria (Salton, 1989; Page et al., 1998).</Paragraph>
    <Paragraph position="1"> Some QA systems on the Web use other techniques: candidate answers are ranked according to a score which takes into account lexical relations between questions and answers, semantic categories of concepts, distance between words, etc. (Moldovan et al., 2003), (Narayanan and Harabagiu, 2004), (Radev and McKeown, 1998).</Paragraph>
    <Paragraph position="2"> Recently, advanced QA systems defined relationships (equivalence, contradiction, ...) between Web page extracts or texts containing possible answers in order to combine them and to produce a single answer (Radev and McKeown, 1998), (Harabagiu and Lacatusu, 2004), (Webber et al., 2002).</Paragraph>
    <Paragraph position="3"> Most systems provide the user with either a set of potential answers (ranked or not), or the &amp;quot;best&amp;quot; answer according to some relevance criteria. They do not provide answers which take into account information from a set of candidate answers or answer inconsistencies. As for logical approaches used for database query, they are based on majority approach or on source reliability. But, contrary to the assumption of (Motro et al., 2004), we noted that reliability information (information about the author, date of Web pages, ...) is rather difficult to obtain, so we assume that all Web pages are equally reliable.</Paragraph>
    <Section position="1" start_page="0" end_page="103" type="sub_section">
      <SectionTitle>
1.2 Motivations and goals
</SectionTitle>
      <Paragraph position="0"> Our framework is advanced QA systems over open domains. Our main goals are to model and to evaluate a system which, from a factoid question in natural language (in French), selects a set of candidate answers on the Web and generates cooperative answers in natural language. Our challenge is (1) to generate a synthetic answer instead of a list of potential answers (in order to avoid providing the user with too much information), and (2) to generate relevant comments which explain the variety of answers extracted from the Web (in order to avoid misleading the user) (Grice, 1975). In a cooperative perspective, we propose an approach for answer generation which uses answer integration. When several possible answers are extracted from the Web, the goal is to define a coherent core  from candidate answers and to generate a cooperative answer, i.e. an answer with explanations. In this paper, we focus on the integration of numerical data in order to generate natural language cooperative answers to numerical questions. We first present some motivational problems for the generation of numerical answers in a QA system.</Paragraph>
      <Paragraph position="1"> Then, we present the content determination and realization processes. Finally, we give some elements of evaluation of our system outputs, with respect to end-users.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>