<?xml version="1.0" standalone="yes"?>
<Paper uid="I05-2030">
  <Title>Opinion Extraction Using a Learning-Based Anaphora Resolution Technique</Title>
  <Section position="3" start_page="173" end_page="173" type="intro">
    <SectionTitle>
2 Related work
</SectionTitle>
    <Paragraph position="0"> In this section, we discuss previous approaches to the opinion extraction problem. In the pattern-based approach (Murano and Sato, 2003; Tateishi et al., 2001), pre-defined extraction patterns and a list of evaluative expressions are used. These extraction patterns and the list of evaluation expressions need to be manually created. However, as is the case in information extraction, manual construction of rules may require considerable cost to provide sufficient coverage and accuracy.</Paragraph>
    <Paragraph position="1"> Hu and Liu (2004) attempt to extract the attributes of target products on which customers have expressed their opinions using association mining, and to determine whether the opinions are positive or negative. Their aim is quite similar to our aim, however, our work differs from theirs in that they do not identify the value corresponding to an attribute. Their aim is to extract the attributes and their semantic orientations.</Paragraph>
    <Paragraph position="2"> Taking the semantic parsing-based approach, Kanayama and Nasukawa (2004) apply the idea of transfer-based machine translation to the extraction of attribute-value pairs. They regard the extraction task as translation from a text to a sentiment unit which consists of a sentiment value, a predicate, and its arguments. Their idea is to replace the translation patterns and bilingual lexicons with sentiment expression patterns and a lexicon that specifies the polarity of expressions. Their method first analyzes the predicate-argument structure of a given input sentence making use of the sentence analysis component of an existing machine translation engine, and then extracts a sentiment unit from it, if any, using the transfer component.</Paragraph>
    <Paragraph position="3"> One important problem the semantic parsing approach encounters is that opinion expressions often appear with anaphoric expressions and ellipses, which need to be resolved to accomplish the opinion extraction task. Our investigation of an opinion-tagged Japanese corpus (described below) showed that 30% of the attribute-value pairs we found did not have a direct syntactic dependency relation within the sentence, mostly due to ellipsis. For example  (The design is weird, but I like it.) This type of case accounted for 46 out of 100 pairs that did not have direct dependency relations. To analyze predicate argument structure robustly, we have to solve this problem. In the next section, we discuss the similarity between the anaphora resolution task and the opinion extraction task and propose to apply to opinion extraction a method used for anaphora resolution.</Paragraph>
  </Section>
class="xml-element"></Paper>