Abstract: Sentiment classification aims to distinguish the expressed sentiment categories by the text, such as positive vs. negative and agree vs. disagree. We used a opinion lexicon, together with a small scale of emotion key words to con- duct sentiment classification with unlabeled data. Specifically, a document word bipartite graph was builts, and then the opinion words and emotion words were served as labeled points while the documents were regarded as unlabeled points in the graph. Label propagation algorithm was used to propagate the label information of the words to the documents. Finally, the high confident automatically-labeled samples were used as training data for sentiment classification through collaborative learning method. Experimental results demonstrate that our approach achieves a good performance for sen- timent classification across multiple domains.

Key words: Emotion words, Sentiment words, Bipartite graph, Label propagation algorithm, Collaborative learning