Aspect-level sentiment analysis based on hierarchical dual-channel graph convolutional network

The attention mechanism-based methods only learn the relevant features of aspect words by concatenating graph convolutional networks and attention operations, without establishing continuous interaction between aspect words and feature representations, resulting in the model's inability to focus on global information. In addition, the model has different sensitivities to syntactic information and semantic information, and it is difficult to aggregate the semantic information of nodes and complex syntactic structure features at the same time using a single dependency relationship, which is not conducive to the final classification of the model. Aiming at the above problems, this paper proposes a dual-channel hierarchical graph convolutional network model, which embeds hierarchical attention into each layer of graph convolutional network to fully aggregate the global features of text information. Among them, emotional knowledge and dependency types are used to supplement the original syntactic dependency graph weighted graph convolutional network to distinguish the importance of different dependency relationships. At the same time, the grammatical structure information is supplemented by contextual semantics to further improve the performance of the aspect sentiment analysis task. The experimental results show that the accuracy and macro F1 score are better than the benchmark model. It can be seen that using the hierarchically aggregated features helps to model semantic and syntactic information.