<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-1641">
  <Title>Sentiment Retrieval using Generative Models</Title>
  <Section position="19" start_page="352" end_page="352" type="ackno">
    <SectionTitle>
DC
BE CUA0BDBNBDCV or
</SectionTitle>
    <Paragraph position="0"> as a set of sentiment seed words D5 D7. For this purpose, we combine sentiment relevance models and topic relevance models, considering the topic dependence of the sentiment. In our experiments, our model worked significantly better than standard language modeling approaches, both when using D5DC and D5D7, and with both manual and automatic annotation of the fragments expressing sentiments in text. With D5 D7 and automatic annotation, our model still worked significantly better than the standard approaches; however, the per- null ment over rmtf where D4 BO BCBMBCBH with the two-sided Wilcoxon signed-rank test.</Paragraph>
    <Paragraph position="1"> formance did not reach that achieved with other settings. We believe the performance can be improved with larger-scale data.</Paragraph>
    <Paragraph position="2"> We experimented to find sentences that were relevant to a given topic and were appropriate to a given sentiment; however, our models can also be applied to textual chunks of any length, such as at document level or passage level. Our model can be easily extended to opinion retrieval, if the opinion retrieval is defined as retrieving sentences or documents that contain either positive or negative sentiments. This issue is worth pursuing in future work. Approaches considering polarity strength or continuous values for the polarity specification, rather than using CUA0BDBNBDCV, can also be considered in future work.</Paragraph>
  </Section>
class="xml-element"></Paper>