关键词: 方面级别情感分析, 图神经网络, 依赖类型, 依赖树距离, 有监督对比损失

Abstract: In the field of aspect based sentiment analysis, existing dual channel models failed to comprehensively consider the different importance levels between grammar nodes, the additional noise introduced by global attention mechanism, and the existence of correlations between similar features. To address the above issues, a dual channel graph convolutional model enhanced by dependency type and dependency distance was proposed. Firstly, dependency types are introduced to measure the importance of neighbourhood nodes; Secondly, mask matrixes based on the dependency tree distance were constructed to filter unrelated noise; Finally, a supervised contrastive loss was introduced to facilitate the model to learn correlations between features that in the same kinds. The experimental results on multiple datasets showed that compared to baseline models, our model gains a better accurancy and macro_f1.

Key words: Aspect-based sentiment analysis, Graph convolutional network, Dependency type, Dependency tree distance, Supervised contrastive loss