Supervised Relation Extraction Based on Labels Contrastive Learning

Relation extraction is an important task in natural language processing. Existing relation extraction tasks usually use data augmentation to construct positive and negative samples for contrastive learning training. And previous works lack the accurate selection of positive and negative samples and the full use of relation labels. To overcome these defects, for supervised relational extraction, we propose a Labels Contrastive Learning framework (LabelsCL) that considers both global and local perspectives, which makes full use of labels for design. Specifically, from a global perspective, we use label and cosine similarity to select the most easily misclassified samples as positive and negative examples for training. From a local perspective, positive and negative samples are selected according to batch and label, and this way of sample selection increases the randomness of sample selection. The contrastive learning framework LabelsCL proposed by us makes full use of labels and synthesizes two perspectives to construct contrastive learning LabelsCL, which increases the tolerance of the framework. Finally, the experimental results demonstrate the effectiveness of our proposed framework, which is significantly improved on the Semeval-2010 Task 8 dataset compared with the baselines.