<?xml version="1.0" standalone="yes"?>
<Paper uid="P05-2008">
  <Title>Using Emoticons to reduce Dependency in Machine Learning Techniques for Sentiment Classification</Title>
  <Section position="6" start_page="47" end_page="47" type="concl">
    <SectionTitle>
5 Conclusions and Future Work
</SectionTitle>
    <Paragraph position="0"> This paper has demonstrated that dependency in sentiment classification can take the form of domain, topic, temporal and language style. One might suppose that dependency is occurring because classifiers are learning the semantic sentiment of texts rather than the general sentiment of language used.</Paragraph>
    <Paragraph position="1"> That is, the classifiers could be learning authors' sentiment towards named entities (e.g. actors, directors, companies, etc.). However, this does not seem to be the case. In a small experiment, we part-of-speech tagged the Polarity 2004 dataset and automatically replaced proper nouns with placeholders.</Paragraph>
    <Paragraph position="2"> Retraining on this modified text did not significantly affect performance.</Paragraph>
    <Paragraph position="3"> But it may be that something more subtle is happening. Possibly, the classifiers are learning the words associated with the semantic sentiment of entities. For example, suppose that there has been a well-received movie about mountaineering. During this movie, there is a particularly stirring scene involving an ice-axe and most of the reviewers mention this scene. During training, the word 'ice-axe' would become associated with a positive sentiment, whereas one would suppose that this word does not in general express any kind of sentiment.</Paragraph>
    <Paragraph position="4"> In future work we will perform further tests to determine the nature of dependency in machine learning techniques for sentiment classification. One way of evaluating the 'ice-axe' effect could be to build a 'pseudo-ontology' of the movie reviews -- a map of the sentiment-bearing relations that would enable the analysis of the dependencies created by the training process. Other extensions of this work are to collect more text marked-up with emoticons, and to experiment with techniques to automatically remove noisy examples from the training data.</Paragraph>
  </Section>
class="xml-element"></Paper>