Prompted representation joint contrastive learning for aspect-based sentiment analysis

Xuefeng Shi, Min Hu, Fuji Ren, Piao Shi

Abstract: Highlights•A novel ABSA model PRCL-GCN is proposed to generate task-oriented contextual representation and construct the relation between the normal graph and the affective graph, which enhances the effectiveness and robustness of the ABSA model.•We design the manual task-oriented prompt template to guide the PLM’s fine-tuning and inject the external affective knowledge into the NDG, which strengthens the sentiment representation in the achieved feature information.•A mutual interaction mechanism is designed based on contrastive learning to effectively reduce the gap between the normal graph and affective knowledge enable graph, and it achieves the truly essential sentiment information from the review for polarity prediction.•Extensive experiments on five benchmark datasets demonstrate that PRCL-GCN performs better than the competitive baselines in terms of accuracy and F1 score.