A neural-based re-ranking model for Chinese named entity recognition

Abstract

Chinese named entity recognition (CNER) is different from English named entity recognition (ENER). There is no specific delimiter in Chinese text to determine the words in a sentence. Besides, the combination of Chinese text has a strong arbitrariness. These special cases usually bring more errors to the Chinese NER (CNER). We propose a re-ranking model based on BILSTM network and without using any other auxiliary methods. Our approach uses N-best generalised label sequences that are produced by baseline model as input and feeds them into our re-ranking model for modelling the context within the generalised sequences. The optimal output sequence is obtained by comprehensively considering the result of baseline model and re-ranking model. Experimental results show that our model achieves better F1-score on Bakeoff-3 MSRA corpus than the best previous experimental results, which yields a 0.97% improvement on F1-score over our neural baseline model and a 0.22% improvement over the state-of-the-art CNER model.