BTDM: A Bi-Directional Translating Decoding Model-Based Relational Triple Extraction

The goal of relational triple extraction is to extract knowledge-rich triples from unstructured text. Although the previous methods obtain considerable performance, there are still some problems, such as error propagation, overlapping problem, and suboptimal subject–object alignment. To address shortcomings above, in this paper, we decompose task into three subtasks a fresh perspective: entity extraction, alignment relation judgement, well propose novel bi-directional translating decoding model (BTDM). Specifically, bidirectional structure designed perform alignment, which decodes pairs both forward backward extraction. effectively mitigates propagation problem aligns pairs. approach handles problem. Finally, (entity pair, relation) bipartite graph achieve practical relationship judgement. Experiments show that our outperforms achieves state-of-the-art performance on NYT WebNLG. We achieved F1-scores 92.7% 93.8% two datasets. Meanwhile, various complementary experiments complex scenarios, demonstrates consistent gain scenarios.