Generate it. A simple method for End-to-End Relation Extraction

Abstract

End-to-end Relation Extraction (RE) is a fundamental problem of information extraction, which includes two tasks: identifying named entities from text and classifying relations between entities. In this work, we propose a simple but effective method to extract entities and relations from text jointly by designing the target output of a BART-based generative model for Named Entity Recognition (NER) without changing its architecture. Compared to existing methods on ChEMU, our method performs better on RE and produces comparable results on NER. Our experimental results also demonstrate that the generative model designed for a single task is capable of joint learning.

Tasks

Reproductions