A Bi-directional Boundary Matching Model for Nested Named Entity Recognition

Span classification for recognizing nested named entities enumerates and verifies all entity spans in a sentence. This strategy has the advantage of resolving nested structures and makes full use of token features in a span. However, it suffers from serious data imbalance and computational complexity problems. In this paper, we propose a span free model to support nested entity recognition, named as bi-directional boundary matching model, where each token in a sentence is used as the centrality of a named entity. Then, a regression operation is applied to locate the left and right boundaries of a named entity. It is an end-to-end multi-objective learning framework, which simultaneously locates entity boundaries relevant to the entity centrality and predicts their confidence scores to be a named entity. We conducted experiments on ACE2005 English and GENIA. Our experiments show that our proposed method achieves the state-of-the-art performance, meanwhile, reduces the computational complexity considerably.