<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-1003">
  <Title>Language and Computation</Title>
  <Section position="3" start_page="18" end_page="18" type="intro">
    <SectionTitle>
2 Background
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="18" end_page="18" type="sub_section">
      <SectionTitle>
2.1 Using Patterns to Extract Semantic Rela-
</SectionTitle>
      <Paragraph position="0"> tions The work discussed here belongs to a line of research attempting to acquire information about lexical and other semantic relations other than similarity / synonymy by identifying syntactic constructions that are often (but not always!) used to express such relations. The earliest work of this type we are aware of is the work by Hearst (1998) on acquiring information about hyponymy (= IS-A links) by searching for instances of patterns such as NP {, NP}* or other NP (as in, e.g., bruises .... broken bones and other INJURIES). A similar approach was used by Berland and Charniak (1999) and Poesio et al (2002) to extract information about part-of relations using patterns such as the N of the N is ....</Paragraph>
      <Paragraph position="1"> (as in the wheel of the CAR is) and by Girju and Moldovan (2002) and Sanchez-Graillet and Poesio (2004) to extract causal relations. In previous work (Almuhareb and Poesio, 2004) we used this same approach to extract attributes, using the pattern &amp;quot;the * of the C [is|was]&amp;quot; (suggested by, e.g., (Woods, 1975) as a test for 'attributehood') to search for attributes of concept C in the Web, using the Google API. Although the information extracted this way proved a useful addition to our lexical representations from a clustering perspective, from the point of view of lexicon building this approach results in too many false positives, as very few syntactic constructions are used to express exclusively one type of semantic relation. For example, the 'attributes' of deer extracted using the text pattern above include &amp;quot;the majority of the deer,&amp;quot; &amp;quot;the lake of the deer,&amp;quot; and &amp;quot;the picture of the deer.&amp;quot; Girju and Moldovan (2002) addressed the problem of false positives for causal relations by developing WordNet-based filters to remove unlikely candidates. In this work, we developed a semantic filter for attributes based on a linguistic theory of attributes which does not rely on WordNet except as a source of morphological information (see below).</Paragraph>
    </Section>
    <Section position="2" start_page="18" end_page="18" type="sub_section">
      <SectionTitle>
2.2 Two Theories of Attributes
</SectionTitle>
      <Paragraph position="0"> The earliest attempt to classify attributes and other properties of substances we are aware of goes back to Aristotle, e.g., in Categories,  but our classification of attributes was inspired primarily by the work of Pustejovsky (1995) and Guarino (e.g., (1992)). According to Pustejovsky's Generative Lexicon theory (1995), an integral part of a lexical entry is its Qualia Structure, which consists of four 'roles':  the Formal Role, specifying what type of object it is: e.g., in the case of a book, that it has a shape, a color, etc.; the Constitutive Role, specifying the stuff and parts that it consists of (e.g., in the case of a book, that it is made of paper, it has chapters and an index, etc.); the Telic Role, specifying the purpose of the object (e.g., in the case of a book, reading); and the Agentive Role, specifying how the object was created (e.g., in the case of a book, by writing).</Paragraph>
      <Paragraph position="1"> Guarino (1992) argues that there are two types of attributes: relational and non-relational. Relational attributes include qualities such as color and position, and relational social roles such as son and spouse. Non-relational attributes include parts such as wheel and engine. Activities are not viewed as attributes in Guarino's classification.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>