Abstract
The task of aspect-level sentiment analysis is to identify the sentiment polarity of sentences when expressed in different aspects. The attention mechanism-based approach allows for attentional interaction between the target and context, but it only combines sentences from a semantic perspective, overlooking the syntactic information present in the sentences. Although graph convolutional networks are capable of handling syntactic information well, they are still unable to effectively combine semantic and syntactic information. This paper proposes a sentiment-supported graph convolutional network (SSGCN), which first extracts the semantic information of words using aspect-aware attention and self-attention. Then, the grammar mask matrix and graph convolutional network are used to combine semantic and grammatical information. The features are then split into two parts - one part extracts semantic and syntactic information related to aspect words, and the other part extracts features related to sentiment-supportive words. Finally, the results from the two parts are concatenated to effectively combine semantic and syntactic information. Experimental results show that the proposed model outperforms the benchmark models in terms of accuracy and macro F1 values on three public datasets.