Prompted representation joint contrastive learning for aspect-based sentiment analysis

As a fine-grained and challenging subtask in the natural language processing (NLP) community, aspect-based sentiment analysis (ABSA) aims to predict the sentiment polarity towards a given aspect term. In previous ABSA research, most works utilized the pre-trained language model (PLM) as the backbone of their proposed methods, without any specific task-related instructions. Besides, some works focused on learning the dependency information or the external knowledge-enhanced dependency information separately, which lacked the exploitation of the mutual interaction between the normal dependency and knowledge-enhanced dependency. Therefore, we propose a novel ABSA method namely prompted representation joint contrastive learning enhanced graph convolutional networks (PRCL-GCN) to strengthen the robustness of the ABSA model. Specifically, to achieve the task-oriented contextual representation, we design the task-specific prompt template to guide the fine-tuning process of PLM in the ABSA task. And a biaffine attention mechanism is employed to further extract the essential sentiment feature from the prompted representation. Moreover, we introduce the syntax dependency graph as prior knowledge, and construct an affective syntactic dependency graph by injecting the affective knowledge from SenticNet into the graph. Then, we utilize the multi-layer GCNs to process the above two syntactic graphs independently, which aims to learn multi-granularity syntactic features. Subsequently, a novel designed attention variant is leveraged to integrate these syntax features with the guided contextual representation, separately. Eventually, through designing a Kullback–Leibler divergence-based contrastive learning to encourage the model’s learning, we improve the model’s accuracy in modeling contextual representation by integrating the designed dual-ways information. Extensive experiments are conducted on five benchmark datasets, and the outstanding experiment results validate the effectiveness of our proposed model.