Sentiment Classification Using Neural Networks with Sentiment Centroids

Maoquan Wang, Shiyun Chen, Liang He

Abstract: Neural networks (NN) have demonstrated powerful ability to extract text features automatically for sentiment classification in recent years. Although semantic and syntactic features are well studied, global category information has been mostly ignored within the NN based framework. Samples with the same sentiment category should have similar vectors in represent space. Motivated by this, we propose a novel global sentiment centroids based neural framework, which incorporates the sentiment category features. The centroids assist NN to extract discriminative category features from a global perspective. We apply our approach to several real large-scale sentiment-labeled datasets, and the extensive experiments show that our model not only obtains more powerful sentiment feature representations, but also achieves some state-of-the-art results with a simple neural network structure.