Chinese NER by Span-Level Self-Attention
Xiaoyu Dong, Xin Xin, Ping Guo · 2019
In this paper, we investigate how to improve Chinese named entity recognition (NER) by applying self-attention mechanism on span-level semantic representations. Specifically, we propose a model which acquires character representations through pre-trained BERT, then extracts features of each possible character-span through LSTM, estimates the semantic reference value of each span, then explicitly leverages span-level information by performing self-attention calculation among span representations. Experiments on OntoNotes 4.0 dataset have demonstrated that the proposed model achieves 79.97% F1-score, outperforming our baseline methods.