<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-2411">
  <Title>Calculating Semantic Distance between Word Sense Probability Distributions</Title>
  <Section position="3" start_page="0" end_page="0" type="relat">
    <SectionTitle>
2 Related Work
</SectionTitle>
    <Paragraph position="0"> Our method draws on, and extends, earlier work in verb lexical semantics (Resnik, 1993; McCarthy, 2000). For example, Resnik (1993) uses relative entropy (KL divergence) to compare the sense profile over the objects of a verb to the profile over the objects of all verbs, to determine how much that verb differs from &amp;quot;average&amp;quot; in its strength of selection for an object. A drawback of this approach for generalizing to other sense profile comparisons is the assumption in relative entropy of an asymmetry between the two probability distributions.</Paragraph>
    <Paragraph position="1"> Similarly, McCarthy (2000) uses skew divergence (a variant of KL divergence proposed by Lee, 1999) to compare the sense profile of one argument of a verb (e.g., the subject position of the intransitive) to another argument of the same verb (e.g., the object position of the transitive), to determine if the verb participates in an argument alternation involving the two positions. For example, the causative alternation in sentences (1) and (2) illustrates how the subject of the intransitive is the same underlying semantic argument (i.e., the Theme--the argument undergoing the action) as the object of the transitive:  (1) The snow melted.</Paragraph>
    <Paragraph position="2"> (2) The sun melted the snow.</Paragraph>
    <Paragraph position="3">  Because we demonstrate our new SPD measure on the same problem as McCarthy (2000), we provide more detail of her method here, for comparison. The first step is to create the sense profiles for the relevant verb/slot pairs (e.g., the intransitive subject of melt, and the transitive object of melt, if determining whether melt undergoes the causative alternation, as illustrated above). The head nouns are extracted from the syntactic slots to be compared for each verb, yielding the frequency of each noun for a verb/slot pair, which is then used to populate the WordNet hierarchy. McCarthy determines the sense profile of a verb/slot pair using a minimum description length tree cut model over the frequency-populated hierarchy (Li and Abe, 1998). The two profiles for a verb are &amp;quot;aligned&amp;quot; to permit comparison using skew divergence as a probability distance measure Lee (1999). (This step is explained in more detail in the next section, with an example.) The value of the distance measure is compared to a threshold, which determines classification of a verb as causative (the two profiles are similar) or non-causative (the two profiles are dissimilar), leading to best performance of 73% accuracy, on a set of hand-selected verbs. In McCarthy (2000), an error analysis reveals that the best method has more false positives than false negatives--some slots are considered overly similar because the sense profiles are compared at a coarse-grained level, losing fine semantic distinctions. Moreover, as mentioned above, the method can only apply to tree-cuts, which restricts its use to a very narrow range of sense profile comparisons.</Paragraph>
    <Paragraph position="4"> In the next section, we propose an alternative method of comparing sense profiles, which addresses each of the shortcomings of these previous measures.</Paragraph>
  </Section>
class="xml-element"></Paper>