Retrieval Contrastive Learning for Aspect-Level Sentiment Classification

ZhongQuan Jian, Jiajian Li, Qingqiang Wu, Junfeng Yao

Abstract: Highlights•We proposed RCL to enable ALSC models to generate polished representations.•RCL has two modules: obtain retrieval instances and learn common features by CL.•The sparse and dense retrievers are used to obtain high-quality retrieval instances.•The ALSC model can be enhanced greatly by training with RCL.•ABSA-DeBERTa obtains new state-of-the-art results by being trained with RCL.