Aspect-based sentiment analysis with weighted relational convolutional network and complementary learning

Aspect-based sentiment analysis in the sight of fine-grained sentiment detects the sentiment polarities towards given aspects terms. The syntactic information of the dependency tree is utilized in existing methods mainly by modeling the equality of different dependencies, which lacks the ability to capture aspect-related globally syntactic information and ignores the importance of considering contextual information of specific aspects. To alleviate the above problems, we propose WRCN-CL with two tasks: Weighted Relational Convolutional Network (WRCN) and auxiliary task complementary learning (CL); Weighted Relational Convolutional Network dynamically assigns weights to prune the dependency tree by considering positional weights and semantic structure simultaneously; Complementary representation is an auxiliary task for knowledge enhancement by obtaining grammatical information related to specific aspects of words. The experimental results prove the effectiveness of the model, and by observing the results on all datasets its state-of-the-art performance proved.