| Chinese Named Entity Recognition(NER)has been an important challenge in natural language processing because Chinese does not have distinct word boundaries like English,more sophisticated techniques are required for recognition.Researchers have developed a variety of different models for the Chinese NER task,including models based on traditional machine learning methods,as well as models based on deep learning methods.In this paper,based on deep learning and machine learning,the following three aspects of work are done to address the problems and difficulties of Chinese NER:First,to address the problems of word boundary blurring and Out-of-vocabulary(OOV)recognition,we propose an adaptive Transformer-based Chinese named entity recognition model,which uses Adapting Transformer Encoder as a character encoder to solve the problem of unregistered word recognition,and then uses Softlexion method is then used to integrate the repertoire information into the character representation to enhance the character representation and solve the problem of blurred word boundaries.In addition,the Adapting Transformer Encoder is used as the semantic layer encoder for encoding,and lastly,the conditional random field is used for label decoding.Secondly,to address the problem of how to integrate lexical information into characters effectively,a lexicon adapter approach is proposed for named entity recognition.The lexicon adapter uses a bilinear attention mechanism to inject lexical information into the character vector and adopts a star topology Star-Transformer for semantic encoding,which makes the performance and inference speed of the model effectively improved.Finally,to learn the word boundary information more explicitly,the model of enhanced boundary perception is proposed.Based on the adaptive Transformer-based model,a multi-task learning approach is used to add two independent Gated Recurrent Unit neural networks as auxiliary tasks to predict the boundaries of entities.Thus,the effectiveness of Chinese named entity recognition is improved.The experiments of this paper are carried out on three Chinese NER datasets,using a unified named entity recognition performance metric,and are validated against other models using lexical enhancement.The experimental performance demonstrates that the F1 values of all three models proposed in thise paper outperform the baseline model and can effectively improve the Chinese NER performance. |
