Aspect Sentiment Triplet Extraction Based on Graph Convolution Neural Network

The method for sequence annotation has a large search space and is prone to errors in the extraction fragments of aspect terms and opinion items, and the use of external syntactic dependency parsers is susceptible to producing parsing errors. In this paper, we propose syntactic dependency and semantic relations enhancement based on Graph Convolutional Network (SE-GCN) for Aspect Sentiment Triplet Extraction. The dilated convolution with multiple expansion rates is used to obtain the semantic information of word pairs at different distances, and the double affine parser is used to construct a word-pair syntactic dependence tree. The syntactic dependence relationship between aspect terms and opinion items in the sentence is established, and the word-pair syntactic feature information is obtained. Finally, the graph convolutional neural model is used to aggregate the semantic and syntactic information of word pairs, update the node features, and predict the emotional polarity. Through multiple experiments, the SE-GCN model achieves good performance on multiple datasets and proves that the addition of syntactic information improves the accuracy of aspect sentiment triplet prediction.