WRTRe: Weighted relative position transformer for joint entity and relation extraction

Entity and relation extraction is a critical task of information extraction in natural language processing. With fast developments of deep learning, this area has attracted great research attention. In spite of these achievements, however, due to the limited feature extraction ability of previous models, extracting overlapping and multiple relation triplets from a sentence is still an enormous challenge. Aim to this issue, here we propose a sequence-to-sequence method, which includes a weighted relative position Transformer encoder to flexibly capture the semantic relationship between entities. To prove the effectiveness of this suggested method, we conduct experiments on two publicly available datasets NYT24 and NYT29. The experimental results show that the proposed approach outperforms previous methods and achieves state-of-the-art performance. Such a framework may shed novel light into knowledge graph construction under complex situations and its potential applications.