<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-2007">
  <Title>Sydney, July 2006. c(c)2006 Association for Computational Linguistics N Semantic Classes are Harder than Two</Title>
  <Section position="4" start_page="49" end_page="49" type="intro">
    <SectionTitle>
2 Relation to Previous Work
</SectionTitle>
    <Paragraph position="0"> Snow et al. (2005) demonstrated binary classification of hypernyms and non-hypernyms using WordNet (Miller, 1995) as a source of training labels. Using dependency parse tree paths as features, they were able to generalize from WordNet labelings to human labelings.</Paragraph>
    <Paragraph position="1"> Turney et al. (2003) combined features to answer multiple-choice synonym questions from the TOEFL test and verbal analogy questions from the SAT college entrance exam. The multiple-choice questions typically do not consist of multiple closely related terms. A typical example is given by Turney: * hidden:: (a) laughable (c) ancient (b) veiled (d) revealed Note that only (b) and (d) are at all related to the term, so the algorithm only needs to distinguish antonyms from synonyms, not synonyms from say hypernyms.</Paragraph>
    <Paragraph position="2"> We use as input phrase pairs recorded in query logs that web searchers substitute during search sessions. We find much more closely related phrases:  * hidden:: (a) secret (e) hiden (b) hidden camera (f) voyeur (c) hidden cam (g) hide (d) spy  This set contains a context-dependent synonym, topically related verbs and nouns, and a spelling correction. All of these could cooccur on web pages, so simple cooccurrence statistics may not be sufficient to classify each according to the semantic type.</Paragraph>
    <Paragraph position="3"> We show that the techniques used to perform binary semantic classification do not work as well when extended to a full n-way semantic classification. We show that using a variety of features performs better than any feature alone.</Paragraph>
  </Section>
class="xml-element"></Paper>