<?xml version="1.0" standalone="yes"?>
<Paper uid="N03-2022">
  <Title>Semantic Extraction with Wide-Coverage Lexical Resources</Title>
  <Section position="3" start_page="0" end_page="0" type="intro">
    <SectionTitle>
2. Background
</SectionTitle>
    <Paragraph position="0"> Semantic Extraction has become a strong research focus in the last few years. A good example is the work of Gildea and Jurafsky (2002) (GJ). GJ present a comprehensive empirical approach to the problem of semantic role assignment. Their work looked at the problem of assigning semantic roles to text based on a statistical model of the FrameNet1 data. In their work, GJ assume that the frame of interest is determined a-priori for every sentence.</Paragraph>
    <Paragraph position="1"> In the IE community, there has been an ongoing effort to build systems that can automatically generate required pattern sets as well as the extraction relevant lexicon. Jones and Riloff (JR) (1999) describe a bootstrapping approach to the problem of IE pattern extension. They use a small seed lexicon and pattern set, to iteratively generate new patterns and expand their lexicon until they achieve an optimized set of patterns and lexicon.</Paragraph>
    <Paragraph position="2"> In the area of lexicon acquisition, many researchers have employed public knowledge bases such as WordNet in IE systems. Bagga et. al. (1997) and later Harabagiu and Maiorano (HM) (2000) investigated the acquisition of the lexical concept space using WordNet and have applied their methods to the Information Extraction task.</Paragraph>
    <Paragraph position="3"> In this paper, we describe work that blends the semantic labeling approach exemplified by the GJ effort and the bootstrapping approach of JR and HM.</Paragraph>
    <Paragraph position="4"> Our work differs from the previous efforts in the following respects. 1) We used FrameNet annotations as seeds both for patterns and for the extraction lexicon. We expand the seed lexicon using WordNet.</Paragraph>
    <Paragraph position="5"> 2) We built a graphical model for the semantic extraction task, which allows us to integrate automatic frame assignment as part of the extraction.</Paragraph>
  </Section>
class="xml-element"></Paper>