Improving Relation Extraction with Contrastive Learning-Based Named Entity Recognition

Named Entity Recognition (NER) and Relation Extraction (RE) are core tasks in information extraction. Pipeline approaches such as PURE insert special markers around entities to guide relation modeling, achieving strong performance with a clear and flexible structure. However, they often transfer NER information only at the output level, lacking model-level integration. This paper proposes a two-stage pipeline framework that introduces a contrastive learning-based bi-encoder in the NER stage to jointly learn span and type representations. The learned encoders are then transferred to the RE stage: the text encoder improves span representation, while the type encoder provides soft-label type information through weighted fusion with type embeddings. This enriched type representation is injected into the RE model for better relation prediction. Experiments on nested NER datasets show that the proposed method outperforms existing pipeline baselines, and ablation studies confirm the effectiveness of model-level knowledge transfer.