Reviewing Labels: Label Graph Network with Top-k Prediction Set for Relation Extraction

The typical way for relation extraction is fine-tuning large pre-trained language models on task-specific datasets, then selecting the label with highest probability of output distribution as final prediction. However, usage Top-k prediction set a given sample commonly overlooked. In this paper, we first reveal that contains useful information predicting correct label. To effectively utilizes set, propose Label Graph Network Prediction Set, termed KLG. Specifically, sample, build graph to review candidate labels in and learn connections between them. We also design dynamic k selection mechanism more powerful discriminative representation. Our experiments show KLG achieves best performances three datasets. Moreover, observe thatKLG effective dealing long-tailed classes.