Latent sentiment representation for sentiment feature selection

Sentiment feature selection (SFS) refers to the task of automatically identifying whether a feature contributes to sentiment classification. Most existing researches do not make a distinction between sentiment classification and topical text classification. Actually, the former commonly depends more on features conveying sentiments while the latter depends on features with strong class distinguish-ability. Therefore, traditional topical feature selection approaches might not be applicable to SFS. In this paper, we propose a novel matrix factorization model for SFS. Our model exploits the sentiment labels of documents to predict words´ sentiment distinguish-ability. Our experiments show that the extracted features are highly accurate and significantly improve the performance in sentiment classification.