Chinese Aspect-Level Sentiment Analysis Based on Graph Convolutional Networks with Syntactic Distance

Aspect-level sentiment analysis tasks aimed at identifying the sentiment polarity of multiple specific aspects of a sentence. Currently most researches based on syntactic analysis have ignored the effect of different relationship types in the dependency tree to the classification results, and most of them are unable to distinguish the importance of node features to the target aspect items. To address these problems, we propose a graph convolutional network model (RD-GCN) that combines relational types and syntactic distances, where the model mainly consists of a Dependency Graph Convolutional Network (DepGCN), a Syntactic Distance Graph Convolutional Network (DisGCN), and an Affine Interaction Module. The DepGCN imports relationship type and dependency probability matrix in the feature transfer process so as to abundantly extract the structural information between the nodes; The DisGCN obtains the importance scores of each word in the sentence according to the syntactic distance matrix from the dependency tree, and setting the distance thresholds to reduce the weight proportion of the long-distance features. The RD-GCN model firstly uses the BiLSTM network as an embedding layer to obtain the initial vectors of the text, then inputs them into two different graph convolutional networks with different syntactic analyses to extract the node features respectively, finally, the affine interaction layer is used as a bridge to make the two different node features interactively learning, to improve the efficiency of the model. The experimental results show that this model on the NLP&CC2012 Chinese dataset improves the Acc and F1 values to 84.36% and 82.58% compared with the ATAE-LSTM and BiGCN comparison models. The Acc and F1 values on the two Chinese datasets of SemEval-2016 task 5 reached 83.60%, 71.99%, and 84.62%, 72.71%, respectively.