Dual Knowledge Aware Graph Convolutional Networks Over Aspect Based Sentiment Analysis
Haoying Si, Jianxia Chen, Lei Mao, Liang Xiao, Haitao Gan, Zhina Song · 2024
Recently, graph convolutional networks (GCNs) have become effective approaches to capture the semantic information of sentences in aspect-based sentiment analysis (ABSA) tasks. Nonetheless, these approaches still face challenges due to syntactic and contextual dependency information from graphs. Therefore, we propose a novel ABSA model by constructing a Dual Knowledge aware Graph Convolutional Networks (DKGCN). First, we obtain the attention score of sentences and aspect words, respectively, via a context-based attention mechanism, to enhance the accuracy of contextual words related to aspect words. Second, we propose a novel syntactic knowledge-based GCN via dependency trees to reduce the connections of words that are not related to the aspect words. Afterward, we propose dual graphs based on context knowledge, consisting of the aspect topic graph and the aspectinfer graph, to achieve contextual knowledge representation of critical aspect words more effectively. In particular, the aspect-topic graph represents relationships between specific aspects and their contextual words according to syntactical dependencies. Moreover, the aspect-infer graph is constructed by an aspect word and other related aspect words inferred by their relationships based on the aspect-topic graph. Finally, we feed dual graphs into the GCNs and the interactive network to extract the fused feature information. Extensive experiments demonstrate that the proposed DKGCN model surpasses advanced baselines in aspect-level sentiment classification.