Aspect Root-Oriented Latent Structures with Graph Convolutional Networks for Aspect Sentiment Classification

Sentiment Analysis based on Aspects (ABSA) is aim to determine the degree of polarity of a set of views. In recent work, dependency trees were combined with graph neural networks (GNN). However, dependency trees can be errors. In order to use syntactic information more effectively, we propose that relies on the root node of the aspect tree and extracting potential syntactic trees from language model (PTM). Hence, our proposed model (RLGCN) can compensate for the inaccuracy of the ordinary dependency trees. The proposed method is proven effective in four datasets.