Research On Chinese Named Entity Recognition Based On Deep Learning

Named entity recognition tasks are processes that identify and label named entities in text.They are a fundamental and important task in natural language processing tasks,such as in machine translation,automatic question answering systems,and information retrieval systems.Chinese named entity recognition task has difficulty in word segmentation,ambiguity,word nesting and structural form complexity compared to English named entity recognition task.It has always been a research hotspot and difficulty in text processing problems.The data set in this article comes from public news text data in the network.We collected and cleaned the text data and used the undersampling and oversampling methods to solve the problem of uneven label classification.The experiment used a BIO format tag to identify five types of named entities and labeled 11 tags.In order to improve the accuracy of the model and the flexibility of the model design,the Bi-GRU-Attention model is proposed based on the RNN-CRF framework.We use a more concise GRU unit of the neural network to simplify the computational complexity,and propose a mechanism that combines attention,improves the disorder of the label,and makes the neural network structure more consistent.This model was designed and tested:Bi-GRU-Attention model proposed in this paper was compared and optimized based on the traditional LSTM-CRF model.The experimental data shows that the training time of the model is shorter than that of the basic model experiment,and the F1 value of the identified five types of named entities is increased by 0.3%.We have also proposed an improved ELMo portable model to address a small number of annotated data problems and to accommodate a variety of needs.In the ELMo model,a straight-through structure is designed to solve the problem of deep convolutional neural network degradation.The word vector is obtained using the improved ELMo model pre-training model and ported to the target task.In the ELMo model applied to the named entity task model,the link layer and the output adjustment layer are designed.The link layer uses the mapping function to solve the problem of inconsistent vector lengths between the pre-trained network and the functional network.The output adjustment layer uses the position information of the word in the word to solve the label correspondence between the output word vector and the word.The experimental results show that the improved ELMo portable model can be transplanted to the named entity recognition task at a small cost,and the F1 value of the control experiment is increased by 1.18%.