Improving Chinese Sentence Polarity Classification via Opinion Paraphrasing

While substantial studies have been achieved on sentiment polarity classification to date, lacking enough opinion-annotated corpora for reliable t rain ing is still a challenge. In this paper we propose to improve a supported vector machines based polarity classifier by enriching both training data and test data via opinion paraphrasing. In particular, we first extract an equivalent set of attribute-evaluation pairs from the training data and then exploit it to generate opinion para-phrases in order to expand the training corpus or enrich opinionated sentences for polarity classification. We tested our system over two sets of online product reviews in car and mo-bilephone domains. The experimental results show that using opinion paraphrases results in significant performance improvement in polarity classification.