<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-3220">
  <Title>Verb Sense and Subcategorization: Using Joint Inference to Improve Performance on Complementary Tasks</Title>
  <Section position="3" start_page="0" end_page="0" type="intro">
    <SectionTitle>
2 Tasks and Data Sets
</SectionTitle>
    <Paragraph position="0"> We evaluate our system on both the WSD task and the verb SCF determination task. We describe each task in turn.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
2.1 Word Sense Disambiguation
</SectionTitle>
      <Paragraph position="0"> We used as our sense-annotated corpus the data sets from the English lexical sample portion of the Senseval-2 word sense disambiguation competition (Kilgarriff and Rosenzweig, 2000). This data set contains multiple instances of 73 different English word types, divided into training and testing examples. Each word type is marked for part of speech, so that the sense disambiguation task does not need to distinguish between senses that have different parts of speech. We selected from this data set all 29 words that were marked as verbs.</Paragraph>
      <Paragraph position="1"> Each example consists of a marked occurrence of the target word in approximately 100 words of surrounding context. The correct sense of the word, marked by human annotators, is also given. Each instance is labeled with a sense corresponding to a synset from WordNet (Miller, 1995). The number of senses per word varies enormously: some words have more than 30 senses, while others have five 2A portion of the Brown corpus has been used both in the construction of the SemCor word sense database and in the construction of the Penn Treebank, but coverage is very low, especially for sense markings, and the individual sentences have not to our knowledge been explicitly aligned.</Paragraph>
      <Paragraph position="2"> or fewer. These &amp;quot;fine-grained&amp;quot; senses are also partitioned into a smaller number of &amp;quot;coarse-grained&amp;quot; senses, and systems are evaluated according to both metrics. The number of training and testing examples per word varies from tens to nearly a thousand.</Paragraph>
      <Paragraph position="3"> We used the same train/test division as in Senseval2, so that our reported accuracy numbers are directly comparable with those of other Senseval-2 submissions, as given in Table 1.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
2.2 Verb Subcategorization
</SectionTitle>
      <Paragraph position="0"> We use as our SCF-annotated corpus sentences drawn from the Wall Street Journal section of the Penn Treebank. For each target verb we select sentences containing a form of the verb (tagged as a verb) with length less than 40 words. We select training examples from sections 2 through 21, and test examples from all other sections.3 There are many conceivable ways to partition the set of possible verb argument combinations into SCFs. One possible approach would be to use as the SCF representation the raw sequence of constituents occurring in the verb phrase. This is certainly an unbiased representation, but as there are many thousands of rewrites for VP in the Penn Treebank, data sparsity would present a significant problem. In addition, many of the variants do not contain useful information for our task: for example, we wouldn't expect to get much value from knowing about the presence or absence of an adverb in the phrase. Instead, we chose to use a small number of linguistically motivated SCFs which form a partition over the large space of possible verb arguments.</Paragraph>
      <Paragraph position="1"> We chose as a starting point the SCF partition specified in Roland (2001). These SCFs are defined declaratively using a set of tgrep expressions that match appropriate verb phrases.4 We made significant modifications to the set of SCFs, and also simplified the tgrep expressions used to match them.</Paragraph>
      <Paragraph position="2"> One difference from Roland's SCF set is that we analyze verb particles as arguments, so that several SCFs differ only in the existence of a particle. This is motivated by the fact that the particle is a syntactic feature that provides strong evidence about the verb sense. One might argue that the presence of a particle should be considered a lexical feature modeled independently from the SCF, but the distinction is blurry, and we have instead combined the variables in favor of model simplicity. A second difference is  NP PP Trans. with prep. phrase VPing Gerundive verb phrase NP VPing Perceptual complement VPto Intrans. w/ infinitival VP NP VPto Trans. w/ infinitival VP S for to Intrans. w/ for PP and infin. VP NP SBAR Trans. w/ finite clause NP NP Ditransitive PRT Particle and no args.</Paragraph>
      <Paragraph position="3"> NP PRT Transitive w/ particle PP PRT Intrans. w/ PP and particle VP PRT Intrans. w/ VP and particle SBAR PRT Intrans. w/ fin. clause and part.</Paragraph>
      <Paragraph position="4"> Other None of the above  that unlike Roland, we do not put passive verb constructions in a separate &amp;quot;passive&amp;quot; SCF, but instead we undo the passivization and put them in the underlying category. Although one might expect that passivization itself is a weak indicator of sense, we believe that the underlying SCF is more useful. Our final set of SCFs is shown in Table 3.</Paragraph>
      <Paragraph position="5"> Given a sentence annotated with a syntactic parse, the SCF of the target verb can be computed by attempting to match each of the SCF-specific tgrep expressions with the verb phrase containing the target verb. Unlike those given by Roland, our tgrep expressions are not designed to be mutually exclusive; instead we determine verb SCF by attempting matches in a prescribed sequence, using &amp;quot;if-thenelse&amp;quot; logic.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>