Aspect-level Sentiment Analysis Based on Heterogeneous Spatial-Temporal Graph Convolutional Networks

As a subtask of fine-grained sentiment analysis, aspect-level sentiment analysis aims to predict the sentiment polarity of aspect words. In recent years, more and more research achievements had been made in this field, but most of the existing methods basically only considered the unified modeling of a paragraph, and the lack of correlation between sentences. The recognition and analysis of emotion not only depended on the content of the sentence itself, but also depended on the emotional fluctuations of the sentence over time. To solve this problem, this paper proposed an aspect-level sentiment analysis model based on Heterogeneous Spatial -Temporal Graph Convolutional Networks (H-STGCN). The model first utilized Bi-LSTM and syntactic dependency tree to model the context, and then constructs each clause as a spatial- temporal graph convolutional network. The attention mechanism calculated the propagation weights, enhancing the feature expression of target aspect words in both time and space dimensions. Experimental results show that the model performs better than the control model on public datasets.