A Chinese named entity recognition method combined with relative position information

Named entity recognition is one of the important tasks of natural language processing, which can help people to select entity information from massive text data. Researchers try to use different methods and improve the recognition effect from different perspectives, including machine learning and deep learning methods, and have made great progress in English datasets. However, in Chinese named entity recognition, it is difficult to recognize entity class because of the complexity of semantic environment and the variety of word formation grammar. Therefore, in order to solve this problem, this paper proposes to use the multi-head attention mechanism of relative position, using the difference of relative position encoding between characters of different positions, to extract the feature of full sentence information, so as to make up for the lack of attention of Lattice-LSTM model to the feature information of full sentence, resulting in the weak ability to recognize complex sentences. Experiments on Chinese Weibo dataset, resume dataset, OntoNotes 4.0 dataset and MSRA dataset verify the model in terms of statement complexity and data volume respectively, and the recognition effect is improved. Finally, we find out a better combination of super parameters, which are further improved on the four datasets.