<?xml version="1.0" standalone="yes"?>
<Paper uid="H05-1045">
  <Title>Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP), pages 355-362, Vancouver, October 2005. c(c)2005 Association for Computational Linguistics Identifying Sources of Opinions with Conditional Random Fields and Extraction Patterns</Title>
  <Section position="8" start_page="360" end_page="360" type="relat">
    <SectionTitle>
7 Related Work
</SectionTitle>
    <Paragraph position="0"> To our knowledge, our research is the first to automatically identify opinion sources using the MPQA opinion annotation scheme. The most closely related work on opinion analysis is Bethard et al.</Paragraph>
    <Paragraph position="1"> (2004), who use machine learning techniques to identify propositional opinions and their holders (sources). However, their work is more limited in scope than ours in several ways. Their work only addresses propositional opinions, which are &amp;quot;localized in the propositional argument&amp;quot; of certain verbs such as &amp;quot;believe&amp;quot; or &amp;quot;realize&amp;quot;. In contrast, our work aims to find sources for all opinions, emotions, and sentiments, including those that are not related to a verb at all. Furthermore, Berthard et al.'s task definition only requires the identification of direct sources, while our task requires the identification of both direct and indirect sources.</Paragraph>
    <Paragraph position="2"> Bethard et al. evaluate their system on manually annotated FrameNet (Baker et al., 1998) and Prop-Bank (Palmer et al., 2005) sentences and achieve 48% recall with 57% precision.</Paragraph>
    <Paragraph position="3"> Our IE pattern learner can be viewed as a cross between AutoSlog (Riloff, 1996a) and AutoSlog-TS (Riloff, 1996b). AutoSlog is a supervised learner that requires annotated training data but does not compute statistics. AutoSlog-TS is a weakly supervised learner that does not require annotated data but generates coarse statistics that measure each pattern's correlation with relevant and irrelevant documents. Consequently, the patterns learned by both AutoSlog and AutoSlog-TS need to be manually reviewed by a person to achieve good accuracy. In contrast, our IE learner, AutoSlog-SE, computes statistics directly from the annotated training data, creating a fully automatic variation of AutoSlog.</Paragraph>
  </Section>
class="xml-element"></Paper>