<?xml version="1.0" standalone="yes"?>
<Paper uid="C02-1078">
  <Title>A Probabilistic Method for Analyzing Japanese Anaphora Integrating Zero Pronoun Detection and Resolution</Title>
  <Section position="5" start_page="25" end_page="25" type="relat">
    <SectionTitle>
4 Related Work
</SectionTitle>
    <Paragraph position="0"> Kim and Ehara (1995) proposed a probabilistic model to resolve subjective zero pronouns for the purpose of Japanese/English machine translation. In their model, the search scope for possible antecedents was limited to the sentence containing zero pronouns. In contrast, our method can resolve zero pronouns in both intra/inter-sentential anaphora types.</Paragraph>
    <Paragraph position="1"> Aone and Bennett (1995) used a decision tree to determine appropriate antecedents for zero pronouns. They focused on proper and definite nouns used in anaphoric expressions as well as zero pronouns. However, their method resolves only anaphors that refer to organization names (e.g., private companies), which are generally easier to resolve than our case.</Paragraph>
    <Paragraph position="2"> Both above existing methods require annotated corpora for statistical modeling, while we used corpora with/without annotations related to anaphoric relations, and thus we can easily obtain large-scale corpora to avoid the data sparseness problem.</Paragraph>
    <Paragraph position="3"> Nakaiwa (2000) used Japanese/English bilingual corpora to identify anaphoric relations of Japanese zero pronouns by comparing J/E sentence pairs. The rationale behind this method is that obligatory cases zero-pronominalized in Japanese are usually expressed in English.</Paragraph>
    <Paragraph position="4"> However, in the case where corresponding English expressions are pronouns and anaphors, their method is not effective. Additionally, bilingual corpora are more expensive to obtain than monolingual corpora used in our method.</Paragraph>
    <Paragraph position="5"> Finally, our method integrates a parameter for zero pronoun detection in computing the certainty score. Thus, we can improve the accuracy of our system by discarding extraneous outputs with a small certainty score.</Paragraph>
  </Section>
class="xml-element"></Paper>