E-strPron: Chinese named entity recognition based on enhanced structure and pronunciation features

Previous researchers founded that CNER(Chinese named entity recognition) models which utilize lexicon, radical and pronunciation features of Chinese characters to enhance Chinese language features can achieve higher performance. However, the strokes and writing order of Chinese characters also include potential semantic information. In this paper, a CNER model E-strPron which utilizes pronunciation, radical, strokes and writing order and lexicon features as Chinese features is proposed. We get the attention of enhanced structure features, pronunciation features and lexicon features with two Cross-Transformers and fuse the mutual attention by soft fusion module. E-strPron is experimented on four well-known CNER benchmark datasets and F1 scores are 81.28% in OntoNotes4.0, 96.06% in MSRA, 96.88% in Resume, 74.5% in Weibo respectively, achieving better performance than baseline models.