<?xml version="1.0" standalone="yes"?>
<Paper uid="N06-1007">
  <Title>Acquisition of Verb Entailment from Text</Title>
  <Section position="3" start_page="0" end_page="49" type="intro">
    <SectionTitle>
2 Previous Work
</SectionTitle>
    <Paragraph position="0"> The task of verb entailment acquisition appears to have much in common with that of paraphrase acquisition (Lin and Pantel, 2001), (Pang et al., 2003), (Szpektor et al., 2004). In both tasks the goal is to discover pairs of related verbs and identify map- null pings between their argument structures. The important distinction is that while in a paraphrase the two verbs are semantically equivalent, entailment is a directional, or asymmetric, relation: one verb entails the other, but the converse does not hold. For example, the verbs buy and purchase paraphrase each other: either of them can substitute its counterpart in most contexts without altering their meaning. The verb buy entails own so that buy can be replaced with own without introducing any contradicting content into the original sentence. Replacing own with buy, however, does convey new meaning.</Paragraph>
    <Paragraph position="1"> To account for the asymmetric character of entailment, a popular approach has been to use lexico-syntactic patterns indicative of entailment. In (Chklovski and Pantel, 2004) different types of semantic relations between verbs are discovered using surface patterns (like &amp;quot;X-ed by Y-ing&amp;quot; for enablement1, which would match &amp;quot;obtained by borrowing&amp;quot;, for example) and assessing the strength of asymmetric relations as mutual information between the two verbs. (Torisawa, 2003) collected pairs of coordinated verbs, i.e. matching patterns like &amp;quot;X-ed and Y-ed&amp;quot;, and then estimated the probability of entailment using corpus counts. (Inui et al., 2003) used a similar approach exploiting causative expressions such as because, though, and so. (Girju, 2003) extracted causal relations between nouns like &amp;quot;Earthquakes generate tsunami&amp;quot; by first using lexico-syntactic patterns to collect relevant data and then using a decision tree classifier to learn the relations. Although these techniques have been shown to achieve high precision, their reliance on surface patterns limits their coverage in that they address only those relations that are regularly made explicit through concrete natural language expressions, and only within sentences.</Paragraph>
    <Paragraph position="2"> The method for noun entailment acquisition by (Geffet and Dagan, 2005) is based on the idea of distributional inclusion, according to which one noun is entailed by the other if the set of occurrence contexts of the former subsumes that of the latter. However, this approach is likely to pick only a particular kind of verb entailment, that of troponymy (such as 1In (Chklovski and Pantel, 2004) enablement is defined to be a relation where one event often, but not necessarily always, gives rise to the other event, which coincides with our definition of entailment (see Section 3).</Paragraph>
    <Paragraph position="3"> march-walk) and overlook pairs where there is little overlap in the occurrence patterns between the two verbs.</Paragraph>
    <Paragraph position="4"> In tasks involving recognition of relations between entities such as Question Answering and Information Extraction, it is crucial to encode the mapping between the argument structures of two verbs. Pattern-matching often imposes restrictions on the syntactic configurations in which the verbs can appear in the corpus: the patterns employed by (Chklovski and Pantel, 2004) and (Torisawa, 2003) derive pairs of only those verbs that have identical argument structures, and often only those that involve a subject and a direct object. The method for discovery of inference rules by (Lin and Pantel, 2001) obtains pairs of verbs with highly varied argument structures, which also do not have to be identical for the two verbs. While the inference rules the method acquires seem to encompass pairs related by entailment, these pairs are not distinguished from paraphrases and the direction of relation in such pairs is not recognized.</Paragraph>
    <Paragraph position="5"> To sum up, a major challenge in entailment acquisition is the need for more generic methods that would cover an unrestricted range of entailment types and learn the mapping between verbs with varied argument structures, eventually yielding resources suitable for robust large-scale applications.</Paragraph>
  </Section>
class="xml-element"></Paper>