Learning to Classify Neutral Examples from Positive and Negative Opinions

Sentiment analysis is a challenging research area due to the rapid increase of subjective texts populating the web. There are several studies which focus on classifying opinions into positive or negative. Corpora are usually labeled with a star-rating scale. However, most of the studies neglect to consider neutral examples. In this paper we study the effect of using neutral sample reviews found in an opinion corpus in order to improve a sentiment polarity classification system. We have performed different experiments using several machine learning algorithms in order to demonstrate the advantage of taking the neutral examples into account. In addition we propose a model to divide neutral samples into positive and negative ones, in order to incorporate this information into the construction of the final opinion polarity classification system. Moreover, we have generated a corpus from Amazon in order to prove the convenience of the system. The results obtained are very promising and encourage us to continue researching along this line and consider neutral examples as relevant information in opinion mining tasks.