A Multi-granularity Contrastive Learning for Distantly Supervised Relation Extraction

Distant supervision, which automates the generation of annotated data, can alleviate the issue of data scarcity in model training. Previous works of distantly supervised relation extraction typically belonged to the bag-level approach, focusing on the context and background information of the entity pair within the instance bags while neglecting the complementarity and interaction between features of varying granularities. We introduce a Multi-Granularity Contrastive Learning for Distantly Supervised Relation Extraction (MGCL) to mitigate the adverse effects of noise in distant supervision. MGCL employs multi-granularity features to construct contrastive learning samples and adopts an asymmetrical contrastive classification strategy during the contrastive phase, aiding the model in understanding textual information from diverse perspectives, thereby obtaining richer and more multidimensional learning signals. In the contrasting process, MGCL also leverages the constraining connections between entities and relations to enrich the semantic background for rare relations. Experiments demonstrate that MGCL surpasses various mainstream baseline models in the relation extraction capability on the public dataset and surpasses the SOTA method in terms of AUC and P@M. This work not only advances the frontiers of distantly supervised relation extraction but also opens new avenues for exploring contrastive learning strategies in noisy environments.