An Improved Method for Chinese Named Entity Recognition Based on MRC

Named Entity Recognition (NER) has been widely using traditional labeling models to identify entities, but its performance is not good in complex entity recognition tasks, such as nested and discontinuous entity recognition. In recent years, paradigm shift between natural language tasks has become one of the hot research topics. This article proposes a new method for improving the performance of NER using the Machine Reading Comprehension (MRC) paradigm. We use the pre-trained language model BERT as an encoder to learn global semantic information, and then combine it with Bi-LSTM to further learn local features. For decoding the head and tail parts of entities in machine reading comprehension, we propose to use GlobalPointer for decoding, which jointly labels the head and tail parts of entities. During training and prediction, entities are used as units, reducing training inconsistencies. The improved method we propose retains the ability of machine reading comprehension to process named entity tasks while improving the ability to correctly decode entities. We conducted extensive experiments on three public datasets: MSRA, OntoNotes 4.0, and CMeEE, achieving improvements of 0.3%, 1.32%, and 1.05% over previous models, respectively.