<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-1112">
  <Title>AnyQ: Answer Set based Information Retrieval System</Title>
  <Section position="3" start_page="5" end_page="6" type="intro">
    <SectionTitle>
4. Retrieval Process
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="5" end_page="6" type="sub_section">
      <SectionTitle>
4.1. Answer Set Search
</SectionTitle>
      <Paragraph position="0"> The main task of answer set search process is capturing a &lt;concept, attribute&gt; pair from natural language query, and mapping them to the knowledge base so that the answer documents can be retrieve.</Paragraph>
      <Paragraph position="1"> User query is represented as natural language so that imply semantic information need, not just single term.</Paragraph>
      <Paragraph position="2"> The query processing distinguishes between the main and additional terms from query. The former covey the essence of the query, reflected &lt;concept, attribute&gt; pairs. The latter help to convey the meaning of the query but can be omitted without changing the essence of the meaning. The secondary terms are useful clue for extracting answer sentence in answer document. Predefined patterns are also important for query processing. As noted earlier, we defined attribute pattern rules for improving accuracy of attribute-based classification. Then we rebuild these patterns as query-attribute pattern  , expecting appeared pattern in interrogative form. Query processing consists of following part: 1) linguistic analyzing, 2) concept focusing, and 3) determining attribute, as illustrated in Figure 3.</Paragraph>
      <Paragraph position="3"> Given a query, what is the problem of angel investment?, we analyze the sentence structure, such as conjunction structure and parallel phrases. We segment complex query into simple sentence. We distinguish the main terms in query by matching the longest term in concept network for focusing concept. To determine attribute of query we first plainly look  for the term in attribute synset, which included title attribute representing all those synonym, i.e.</Paragraph>
      <Paragraph position="4"> Problems is title of set {Warning, danger, abuse, damage}. If not, we classify the question into one of 83 categories, each of mapped to a particular set of query-attribute pattern. Our example query map &lt;angel investment, problem&gt; pair.</Paragraph>
      <Paragraph position="5"> After extracting appropriate &lt;concept, attribute&gt; pair, query expansion is generated, connecting with related concept in concept network. This expansion is based on the assessment of similarity between distances of concept network. The main advantage of connecting related concept is that the user can be traverse concept network through semantic path.</Paragraph>
      <Paragraph position="6"> Thus continuous search feedback can be possible.</Paragraph>
      <Paragraph position="7"> Expanded query map to knowledge base so that the documents corresponding the &lt;concept, attribute&gt; pairs can be retrieve as answer set. The results were ranked using attributed-based classification score in answer set construction processing.</Paragraph>
    </Section>
    <Section position="2" start_page="6" end_page="6" type="sub_section">
      <SectionTitle>
4.2. Result Presentation: Highlighting
Answer Sentence
</SectionTitle>
      <Paragraph position="0"> Finding answer to a natural language question involves not only knowing what the answer is but also where the answer is. The answer set that produced initial searching step is considered to include the candidate answer sentence. For detecting answer sentence, we extract all the possible &lt;concept, attribute&gt; pairs each sentence in answer document.</Paragraph>
      <Paragraph position="1"> The sentence was not include query pairs was discard so that we can get candidate sentences where answer is appeared. Similarly query expansion, candidate sentences calculate score of match additional query words, which is generated in query processing.</Paragraph>
      <Paragraph position="2"> Highest scoring sentence was highlighted including its former and latter sentence. Right side box in Figure 3 shows our retrieval process.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>