Dual BiGRU-CNN-based sentiment classification method combining global and local attention

Traditional sentiment classification methods often ignore the influence of psychological features on classification results. Moreover, most sentiment classification methods fail to effectively integrate the context semantic information of local structural features and mine the interaction between context semantic features and local structural features. On this basis, we proposed a novel dual BiGRU-CNN-based sentiment classification method combining global and local attention. First, in order to integrate psychological information into sentiment classification, the advantage of the language query and word count model (LIWC) on effectively expressing the users’ psychological features is utilized, and a heterogeneous graph attention networks (HAN) and LIWC-based text representation learning method (called HAN_LIWC) are proposed to improve the representation of each document. Then, we propose a new sentiment classification framework (called BCAT) in which the word features with context semantic information and the local structural features are comprehensively considered to improve the sentiment classification accuracy. Finally, we introduce the global attention layer and the local attention layer. In order to utilize the mutual effect between different features, the global attention layer focuses on mining the interaction between word features, the interaction between local structural features, and the interaction between word features and local structural features. Moreover, the local attention layer focuses on mining the contributions of all features to the final classification results. Experimental results on three datasets show that the proposed algorithm has greatly improved classification accuracy compared to typical traditional machine learning-based methods and deep learning-based methods.