Chinese Named Entity Recognition Based on Dynamically Adjusting Feature Weights

Named entity recognition is a basic task in NLP, and it is an important basic tool for many NLP tasks such as information extraction, parsing, question answering system and machine translation. The extraction of sequence features of datasets directly affects the recognition effect of named entities, and only the accumulation of local sequence features cannot capture the long distance dependencies. The extraction of global sequence features improves this problem, but loses some local features. Long entities are nested within short entities and have different entity attributes from short entities, resulting in identification errors. To solve these problems, a Chinese named entity recognition algorithm based on Bert +FL-LGWF+CRF is proposed. In this method, the text is encoded into a word vector matrix by Bert as the input to FL-LGWF (Entity Level-Local And Global Weighted Fusion). FL-LGWF utilizes CNN (Convolutional Neural) to extract the local sequence features of the text vector, and use BISTM (Bidirectional Long Short-Term Memory) to extract contextual global sequence features, and perform dynamic weight fusion on the extracted sequence features. Then the score matrix of the tag is obtained according to the entity attribute level. Finally, the global optimal tag sequence is obtained through the CRF layer. Experimental results show that the proposed Bert +FL-LGWF+CRF model has higher F1 value on both public data sets and self-created data sets.