Sentiment Embedded Semantic Space for More Accurate Sentiment Analysis.

Word embedding is one common word vector representation with improved performance for sentiment analysis task. Most existing methods of learning context-based word embedding are semantic oriented, but they typically fail to capture the sentiment information. This may result in words with similar vectors but with very different sentiment polarities, thus degrading the followed sentiment analysis performance. In this paper, we propose a novel and efficient method to yield the Sentiment Embedded Semantic Space that captures the connection between the sentiment space and the semantic space. The proposed method is based on K-means and CNN. In addition, we develop a more fine-grained sentiment dictionary based on HowNet Dictionary together with the processing dataset. Extensive experiments on benchmark datasets show that the proposed method leads to more accurate sentiment classifier and reduces the task-specific word embedding effort.