A Mixed Graph Convolutional Network with Cross-Distance Syntactic Aware for Aspect-Based Sentiment Analysis
Ruhan Deng, Qinghua Zhang, Man Gao, Meiling Fu, Ye Wang, Guoyin Wang · 2024
Aspect-based sentiment analysis(ABSA) is one of the text classification tasks, aiming to classify sentiment of aspect words in the given text. Since syntactic information of text can be effectively modeled by dependency tree. Meanwhile, local and global information of sentences can be well learned by graph convolutional network(GCN), so dependency tree-based GCN is widely used in ABSA, where the relationships between sentence words can be effectively captured. However, deep GCN suffers from the problem of over-smoothing, which simply means that nodes of different orders are indistinguishable, especially on small scale datasets. And syntactic information is enhanced early in model training in previous studies. In order to address these problems from other perspectives, first, a mix graph convolutional network with cross-distance syntactic aware for ABSA is proposed, to alleviate nodes over-smoothing of GCN problem. Second, Mixhop is combined with GCN, to help the aspectual node to acquire feature information of multi-layer neighbors. The proposed model, not only allows aspect words to fuse the information learned from the shallow-layer neighbors, but also fuses the information from deep-layer neighbors. Third, attention mechanism is used to focus on syntactic information of nodes that are distant from the aspect words. Finally, experiments conducted on widely used public datasets that the proposed model achieves superior results.