A Relation-Attentive 3D Matrix Framework for Relational Triple Extraction

Anonymous

Abstract: Extracting relational triples from unstructured text is crucial for information extraction. Recent methods achieve considerable performance, but due to the insufficient consideration of triple global information, there is an obvious performance gap between triple (E1, R, E2) and E1/R/E2, that is, some extracted entities or relations fail to form a valid relational triple. To break this bottleneck, we propose a relation-attentive 3D matrix framework (RA3D) composed of an encoder module, a fusion module, and a 3D matrix module. Instead of using a 2D table to align the subject and object, we integrate clearly encoded relation information to convert the 2D table into a 3D matrix, so that the entries of the 3D matrix can capture the interaction in subjects, objects, and relations completely. To extract relation and entity information required for the 3D matrix reasonably, we design a transformer-decoder-based fusion module that updates the representation of relations and entities iteratively. Our model achieves state-of-the-art performance with F1 score up to 93.5\% and 94.3\% on two public datasets and delivers consistent performance gain on complex scenarios of overlapping triples.