<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-0805">
  <Title>Exploring Semantic Constraints for Document Retrieval</Title>
  <Section position="3" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> The questions of where and how sophisticated natural language processing techniques can improve traditional term-based information retrieval have been explored for more than a decade. A considerable amount of work has been carried out that seeks to leverage semantic information for improving traditional IR. Early TREC systems such as INQUERY handled both natural language and semi-structured queries and tried to search for constraint expressions for country and time etc. in queries (Croft et al., 1994). Later work, as discussed in (Strzalkowski et al., 1996), has focused on exploiting semantic information at the word level, including various attempts at word-sense disambiguation, e.g., (Voorhees, 1998), or the use of special-purpose terms; other approaches have looked at phrase-level indexing or full-text query expansion. No approaches to date, however, have sought to employ semantic information beyond the word level, such as that expressed by attribute-value (AV) pairs, to improve term-based IR.</Paragraph>
    <Paragraph position="1"> Attribute-value pairs offer an abstraction for instances of many application domains. For example, a person can be represented by a set of attributes such as name, date-of-birth, job title, and home address, and their associated values; a house has a different set of attributes such as address, size, age and material; many product specifications can be mapped directly to AV pairs. AV pairs represent domain specific semantic information for domain instances.</Paragraph>
    <Paragraph position="2"> Using AV pairs as semantic constraints for retrieval is related to some recent developments in areas such as Semantic Web retrieval, XML document retrieval, and the integration of IR and databases. In these areas, structured information is generally assumed. However, there is abundant and rich information that exists in unstructured text only. The goal of this work includes first to explore a method for automatically extracting structured information in the form of AV pairs from text, and then to utilize the AV pairs as semantic constraints for enhancing traditional term-based IR systems.</Paragraph>
    <Paragraph position="3"> The paper is organized as follows. Section 2 describes our method of adding AV annotations to text documents that utilizes a domain model automatically extracted from the Web. Section 3 presents two IR systems using a vector space model and semantic constraints respectively, as well as a system that combines the two. Section 4 describes the data set and topic set for evaluating the IR systems. In Section 5, we compare the performance of the three IR systems, and draw initial conclusions on how NLP techniques can improve traditional IR in specific domains.</Paragraph>
  </Section>
class="xml-element"></Paper>