Abstract: Text sentiment classification is a very important branch of natural language processing.Researchers focus on the accuracy of sentiment classification but ignore the time cost of training and classification.Bag-of-words feature used in most methods for text sentiment classification has high dimension and bad interpretability.To solve the above problems,we presented a multi-granularity text sentiment classification model based on three-way decisions for document-level sentiment classification.With the aid of granular computing,we made a structure of text that contains three levels of granularity－word,sentence and document,and presented a new kind of feature－SSS(sentence-level sentiment strength) feature which represents a document,in which the value of each dimension is the sentence-level sentiment strength.In classification process,we firstly utilized three-way decisions method to divide the objects into three regions.The objects in positive region and negative region are classified into positive class and negative class,respectively.We employed the state-of-the-art classifier－SVM to classify the objects in boundary region.Experimental results show that combining three-way decisions method and SVM can improve the accuracy of classification.The SSS-feature reduces the time-cost of feature extraction and training greatly because of its low dimension.Three-way decisions method can reduce the time-cost of classification,and they can ensure good performance in classification accuracy at the same time.

Key words: Sentiment classification,Three-way decisions,Multi-granularity,SVM