Contrastive Triple Extraction with Generative Transformer

Triple extraction is an essential task in information for natural language processing and knowledge graph construction. In this paper, we revisit the end-to-end triple sequence generation. Since generative may struggle to capture long-term dependencies generate unfaithful triples, introduce a novel model, contrastive with transformer. Specifically, single shared transformer module encoder-decoder-based To faithful results, propose triplet training object. Moreover, two mechanisms further improve model performance (i.e., batch-wise dynamic attention-masking triple-wise calibration). Experimental results on three datasets NYT, WebNLG, MIE) show that our approach achieves better than of baselines.