Word distance assisted dual graph convolutional networks for accurate and fast aspect-level sentiment analysis

Aspect-level sentiment analysis can provide a fine-grain sentiment classification for inferring the sentiment polarity of specific aspects. Graph convolutional network (GCN) becomes increasingly popular because its graph structure can characterize the words' correlation for extracting more sentiment information. However, the word distance is often ignored and cause the cross-misclassification of different aspects. To address the problem, we propose a novel dual GCN structure to take advantage of word distance, syntactic information, and sentiment knowledge in a joint way. The word distance is not only used to enhance the syntactic dependency tree, but also to construct a new graph with semantic knowledge. Then, the two kinds of word distance assisted graphs are fed into two GCNs for further classification. The comprehensive results on two self-collected Chinese datasets (MOOC comments and Douban book reviews) as well as five open-source English datasets, demonstrate that our proposed approach achieves higher classification accuracy than the state-of-the-art methods with up to 1.81x training acceleration.