Sentiment classification using weakly supervised learning techniques

Due to the advanced technologies of Web 2.0, people are participating in and exchanging opinions through social media sites such as Web forums and Weblogs etc., Classification and Analysis of such opinions and sentiment information is potentially important for both service and product providers, users because this analysis is used for making valuable decisions. Sentiment is expressed differently in different domains. Applying a sentiment classifiers trained on source domain does not produce good performance on target domain because words that occur in the train domain might not appear in the test domain. We propose a hybrid model to detect sentiment and topics from text by using weakly supervised learning technique. First we create sentiment sensitive thesaurus using both labeled and unlabeled data from multiple domains. The created thesaurus is then used to classify sentiments from text. This model is highly portable to various domains. This is verified by experimental results from four different domains where the hybrid model even outperforms existing semi-supervised approaches.