Intrinsic Dependency Graph Convolutional Networks for Aspect Level Sentiment Analysis
Ying Hou, Fang'ai Liu, Xuqiang Zhuang, Yuling Zhang, Lin Zhang · 2024
Aspect-based sentiment analysis (ABSA) aims to identify specific aspects within text and determine their associated sentiment polarities. Addressing this intricate challenge, various robust approaches, such as attention mechanisms and Convolutional Neural Networks (CNNs), have been widely adopted. Studies have indicated that the use of Graph Convolutional Networks (GCN) based on semantic dependency trees can yield improved results. Consequently, a myriad of methods have emerged that leverage sentence structure to complete the task. Nonetheless, many of these methods fail to account for directional dependencies between terms and their context as well as the inherent interdependencies among the terms themselves. In our study, we introduce an innovative model, the Intrinsic Dependency Graph Convolutional Network (IDGCN), which refines the integration of directional dependencies within graph convolutions to more accurately represent information in a time-series context. We also present a dependency encoder designed to reinforce the interconnections among contextual elements. Through comprehensive experiments and comparative analyses across various subtasks, we demonstrate the superior performance of the IDGCN. The experimental outcomes from four distinct datasets confirm the efficacy of our approach.