Boundary regression model for joint entity and relation extraction

Abstract

Joint extraction of named entities and their relations has the advantage of avoiding cascading failures caused by falsely recognized named entities. Recent studies have focused on span classification modes to support end-to-end multiobjective learning. However, the enumeration of a large number of inaccurate entity spans creates a serious data imbalance and incurs high computational complexity. In this study, we propose a boundary regression model for joint entity and relation extraction, where a boundary regression mechanism is adopted to learn the offset of a possible named entity relevant to a true named entity. Instead of exhaustively enumerating all possible entity spans, this model receives only a small number of coarse entities with inaccurate boundaries as inputs. It can locate named entities and extract relations between them simultaneously. Experiments demonstrated that our boundary regression model outperforms state-of-the-art models in terms of the F1 score by ＋2.5%, ＋0.4%, ＋2.1%, and ＋1.3% on ADE, ACE05, ACE04, and CoNLL04 benchmark datasets respectively. Analytical experiments further confirmed the effectiveness of our model for refining entity boundaries and learning accurate span representations.