<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-3506">
  <Title>Catching Metaphors</Title>
  <Section position="8" start_page="45" end_page="46" type="relat">
    <SectionTitle>
7 Related Work
</SectionTitle>
    <Paragraph position="0"> Previous work on automated metaphor detection includes Fass (1991), Martin (1990), and Mason (2004). Whereas our aim is to classify unseen sentences as literal or metaphorical, these projects address the related but distinct task of identifying metaphoricalmappings. Allthreeusetheselectional preferences of verbs to identify metaphors. In literal usage, the arguments that fill particular roles of a verb are frequently of a common type. For instance, in the MEDICAL domain, the object of the  to total examples for metaphor (M), literal (L) and total (Total) is shown. The total percent correct for the frame (%Tot), the overall baseline percentage (%OBL), and the verb baseline percentage (%VBL) are also shown. The cumulative performance over all frames is located in the bottom row of the table. verb treat is usually a pathological state. In the FI-NANCE domain, the object of treat is usually an economic problem. This difference in selectional preference suggests metaphorical usage. Furthermore, it suggests a metaphorical mapping between health problems and economic problems.</Paragraph>
    <Paragraph position="1"> The systems described by Fass and Martin exhibit impressive reasoning capabilities such as identifying novel metaphors, distinguishing metaphor from metonymy, and interpreting some metaphorical sentences. But they require hand-coded knowledge bases and thus have limited coverage and are difficult to extend. More similar to our efforts, Mason's CorMet uses a corpus-based approach. In CorMet, domains are characterized by certain key-words which are used to compile domain-specific corpora from the internet. Based on differences in selectional preferences between domains, CorMet seeks to identify metaphorical mappings between concepts in those domains.</Paragraph>
    <Paragraph position="2"> One shortcoming of using syntactic arguments is reflected by CorMet's mistaken identification of a mapping between institutions and liquids. This arises from sentences like The company dissolved and The acid dissolved the compound. Such sentences suggest a mapping between the subjects in the target domain, institutions, and the subjects in source domain, liquids. Using semantic roles avoids this source of noise. This is not to suggest that the syntactic features are unimportant, indeed the selectional preferences determined by CorMet could be used to select which arguments to use for features in our classifier.</Paragraph>
    <Paragraph position="3"> Our approach considers each sentence in isolation. However the distribution of metaphorical usage is not uniform in the WSJ corpus (Martin, 1994),. It is therefore possible that the information about surrounding sentences would be useful in determining whether a usage is metaphorical. CorMet incorporates context in a limited way, computing a confidence rating, based in part upon whether a metaphoric mapping co-occurs with others in a systematic way.</Paragraph>
  </Section>
class="xml-element"></Paper>