Aspect-level sentiment classification with aspect-opinion sentence pattern connection graph convolutional networks

Attempting to identify and determine the sentiment polarity of one or more aspects (i.e., aspect words) in a sentence, Aspect-Level Sentiment Classification (abbreviated as ALSC) is a fine-grained sentiment classification task. Graph convolutional networks on dependency trees are now widely being used in related research to improve the accuracy of ALSC. The key to determining the polarity of aspectual emotions is to find the opinion, i.e., the opinion word, that is, most relevant to the aspectual emotion. However, in the dependency tree, a significant portion of aspect words and opinion words are not directly connected. And long-distance connections can lead to the model not paying enough attention to opinion words and losing information. In order to address this issue, by examining dependency syntactic structure and syntactic knowledge, we propose Aspect-opinion Sentence pattern Connection (ASC) to strengthen sentiment dependency graphs. We then develop the ASC-GCN to efficiently use the strengthened dependencies. Experimental results on four public benchmark datasets indicate that our approach achieves excellent performance on a lightweight model.