Joint Modeling of Entities and Relations Through Hierarchical Analysis of Span Sequences

Relation extraction, a crucial aspect of information extraction, has greatly benefited from recent progress in deep neural networks and pretrained language models. Two dominant approaches emerge: the cascade method and the joint approach. The latter, though ambitious, faces challenges related to entity overlap. A promising alternative is the span-based approach, but it is hindered by limitations in span representation. To overcome these obstacles, we propose an innovative joint extraction model. It relies on multi-level feature analysis at the span representation level, detecting local patterns and sequential dependencies. This approach enables the identification of crucial information at a fine scale, thereby improving extraction accuracy. We trained our model using two datasets, NYT and WEBNLG, and evaluated its effectiveness on the latter. Our approach surpasses other models, achieving superior performance, especially in terms of the F1 score on WEBNLG, underscoring its potential for relation extraction.