A Joint Relational Triple Extraction Model Based on Relational Bidirectional Encoding

Tagging based relational triple extraction methods have recently gained significant research attention. However, most of these methods suffer from the limitation of a unidirectional extraction framework. This framework is overly reliant on the subjects identified by the model in the first stage, leading to an Error Propagation Problem. To conquer this limitation, we devise a relational bidirectional encoding framework that identifying triples and its inverse based on relation and its inverse, then combining the scores of bidirectional triples to obtain factual triples. We use a shared DAE (Denosing Autoencoder) to recognize bidirectional entity pairs, so that the bidirectional features can flow during their respective recognition processes, thereby enhancing and complementing each other. Then we employ a bidirectional biaffine score function to assign every relation to every forward and inverse entity pair. During training, we address the problem of convergence rate inconsistency stemming from shared DAE through introducing a multi-task learning scheme. Our model is experimented on several typical benchmark datasets, and the results demonstrate its effectiveness. It achieves SOTA (state-of-the-art) performance on almost all datasets. Furthermore, our relational bidirectional encoding framework and multi-task learning scheme exhibit excellent adaptability and can improve the precision of other Tagging based factual triple recognition models.