Multi-branch Attention Fusion Network for Aspect Sentiment Classification

For specific aspects (aspect words) of the object described in a given sentence, the research goal of aspect sentiment classification is to predict the sentiment polarity expressed by different aspect words. At present, sentiment classification methods based on attention mechanism and graph neural networks use attention models to capture contextual information related to aspect words, but ignore the global information of the sentence where aspect words are located. The classification accuracy in dealing with complex sentences still needs to be further improved. To solve the above problems, we propose a novel syntax-aware method - Multi-branch Attention Fusion Network (MAFNet) for aspect sentiment classification. Firstly, we use graph attention networks to capture contextual information related to aspect words, while using the mean-pooling layer to capture global information of sentences. Secondly, we propose a novel attention fusion mechanism to integrate contextual information related to aspect words and global information of sentences, which can enrich the semantic relationship between aspect words and opinion words in the sentence. Then, we use orthogonal and differential regularization to constrain the attention weight of contextual information, which can accurately capture the semantic correlation between aspect words and opinion words in the sentence. Finally, we verify the effectiveness of the algorithm on three public datasets. The experimental results prove that our method can correctly establish the semantic relationship between aspect words and opinion words, which can achieve better classification accuracy than traditional methods, attention-based methods, and graph neural network-based methods.