<?xml version="1.0" standalone="yes"?>
<Paper uid="E06-1042">
  <Title>A Clustering Approach for the Nearly Unsupervised Recognition of Nonliteral Language[?]</Title>
  <Section position="3" start_page="329" end_page="329" type="intro">
    <SectionTitle>
2 Previous Work
</SectionTitle>
    <Paragraph position="0"> The foundations of TroFi lie in a rich collection of metaphor and metonymy processing systems: everything from hand-coded rule-based systems to statistical systems trained on large corpora. Rule-based systems - some using a type of interlingua (Russell, 1976); others using complicated networks and hierarchies often referred to as metaphor maps (e.g. (Fass, 1997; Martin, 1990; Martin, 1992) - must be largely hand-coded and generally work well on an enumerable set of metaphors or in limited domains. Dictionary-based systems use existing machine-readable dictionaries and path lengths between words as one of their primary sources for metaphor processing information (e.g. (Dolan, 1995)). Corpus-based systems primarily extract or learn the necessary metaphor-processing information from large corpora, thus avoiding the need for manual annotation or metaphor-map construction. Examples of suchsystems canbefound in(Murata et. al., 2000; Nissim&amp;Markert, 2003; Mason, 2004). Thework on supervised metonymy resolution by Nissim &amp; Markert and the work on conceptual metaphors by Mason come closest to what we are trying to do with TroFi.</Paragraph>
    <Paragraph position="1"> Nissim &amp; Markert (2003) approach metonymy resolution with machine learning methods, &amp;quot;which [exploit] the similarity between examples of conventional metonymy&amp;quot; ((Nissim &amp; Markert, 2003), p. 56). They see metonymy resolution as a classification problem between the literal use of a word and a number of pre-defined metonymy types.</Paragraph>
    <Paragraph position="2"> They use similarities between possibly metonymic words (PMWs) and known metonymies as well as context similarities to classify the PMWs. The main difference between the Nissim &amp; Markert algorithm and the TroFi algorithm - besides the fact that Nissim &amp; Markert deal with specific types of metonymy and not a generalized category of nonliteral language - is that Nissim &amp; Markert use a supervised machine learning algorithm, as opposed to the primarily unsupervised algorithm used by TroFi.</Paragraph>
    <Paragraph position="3"> Mason (2004) presents CorMet, &amp;quot;a corpus-based system for discovering metaphorical mappings between concepts&amp;quot; ((Mason, 2004), p. 23). His system finds the selectional restrictions of given verbs in particular domains by statistical means. It then finds metaphorical mappings between domains based on these selectional preferences. By finding semantic differences between the selectional preferences, it can &amp;quot;articulate the higher-order structure of conceptual metaphors&amp;quot; ((Mason, 2004), p. 24), finding mappings like LIQUID-MONEY. Like CorMet, TroFi uses contextual evidence taken from a large corpus and alsousesWordNetasaprimaryknowledge source, but unlike CorMet, TroFi does not use selectional preferences.</Paragraph>
    <Paragraph position="4"> Metaphor processing has even been approached with connectionist systems storing world-knowledge as probabilistic dependencies (Narayanan, 1999).</Paragraph>
  </Section>
class="xml-element"></Paper>