Abstract: Aiming at  the problem of sparse entity relationship type data with unclear feature information when using global pointer networks for entity relationship extraction, as well as the problem of class imbalance and incorrect labeling in the data, we proposed a entity-relation joint extraction model based on  contrastive learning and gradient penalty methods while utilizing an enhanced RoBERTa pre-trained model. Experimental results on the Alibaba Tianchi Chinese medical information processing benchmark CBLUE2.0 dataset show  that this model outperforms the global pointer network, and can more  effectively extract  entity relationship from complex data.

Key words: entity relation extraction, contrastive learning, gradient penalty, RoBERTa pre-trained model, global pointer network