MPGM:Multi-prompt generation model with self-supervised contrastive learning for aspect sentiment triplet extraction

Generative models are widely used in natural language processing and achieve remarkable results in the Aspect Sentiment Triplet Extraction (ASTE) task. Although existing generative methods can effectively identify triplets within sentences, their performance still needs to be improved when dealing with complex sentences containing multi-span terms. The main issues are the insufficient recognition of long-span terms and the lack of comprehensive recognition of complete triplets. To address the limitations of generative models in the ASTE task, a multi-prompt generation model (MPGM) with self-supervised contrastive learning is proposed for aspect sentiment triplet extraction. Efforts are made to enhance the connections between terms and sentiment polarity from various perspectives. Initially, multiple prompt templates are proposed to integrate the generated triplets to mitigate potential errors in individual templates. In addition, a terminological affinity evaluation is designed, which incorporates term information during the training process to enhance the model's ability to recognize relationships between terms. Moreover, the dual-dimensional supervised contrastive learning strategy leverages multiple types of labels to enhance the representations of triplet spans. The enhanced triplet span representations facilitate more precise modeling of the relationships between terms and sentiment polarities. Extensive experiments have validated that the MPGM demonstrates superior performance to existing methods on two public datasets, proving its effectiveness and advancement in addressing the challenges of the ASTE task. Specifically, on the four subsets of the ASTE-DATA-v1 and ASTE-DATA-v2 datasets (14Lap, 14Res, 15Res, 16Res), the F1 scores of the MPGM method are 63.98 %, 76.63 %, 67.48 %, 75.61 % and 65.32 %, 76.80 %, 68.75 %, 75.49 %, respectively.