TDGAT: Triple-Dimensional Graph Attention Networks for Exploring the Optimal Perspective for Aspect-Based Sentiment Analysis

Xuefeng Shi, Weiping Ding, Min Hu, Xin Kang, Fuji Ren

Abstract: Aspect-based sentiment analysis (ABSA) is an attractive and challenging fine-grained subtask in the natural language processing (NLP) community. In prior works, the relevant research has achieved significant progress in leveraging external knowledge to strengthen the sentiment representation in ABSA. However, the prior simple utilization of external knowledge hurts the model’s comprehension of the diversity of sentiment. Therefore, this paper proposed a novel ABSA approach to achieve multi-dimensional understanding of sentiment, namely triple-dimensional graph attention networks (TDGAT). Preliminarily, a task-oriented prompt template is designed to evoke the pre-trained language model (PLM) generation ability, which would provide the essential semantic features for sentiment understanding. To avoid the errors caused by the single-dimension external knowledge, TDGAT constructs triple-dimensional sentiment graphs depending on valence, arousal, and dominance, respectively, which enhance the model’s sentiment comprehension ability greatly. Besides, this work also leverages GATs to evacuate the sentiment relations through exploiting the constructed graphs, which would largely assist to learn an optimal ABSA model. Eventually, to uncover the optimal sentiment exploration perspective, extensive experiments are conducted on five public and available benchmark datasets, and the related results show the proposed TDGAT outperforms the-state-of-art baselines with different sentiment perspectives.

External IDs:dblp:journals/cogcom/ShiDHKR25