The Research On Improvement Of Chinese Named Entity Recognition Method Based On Deep Learning

With the high development of information technology,the speed of text data generation on the Internet has increased dramatically.However,these textual data are hardly used directly for their unstructured form despite the abundant valuable information they have.Information extraction,which can extract structured data from a large number of unstructured ones,is among the research emphases of signal processing.Named Entity Recognition(NER)also has valuable meanings,as a foundations of information extraction technology.NER is a method to recognize named entities as name of people,location and organization from unstructured data.Traditionally,it based on rules and statistics,which lacked versatility,highly dependent on linguistic knowledge and cost much manpower on designing features.In this case,the method based on Deep Learning(DL)is more common.The DL-based NER method required no human intervention in the whole process,and can learn features automatically through Artificial Neural Network(ANN),thus this method gets rid of the dependence on linguistic knowledge.Among the NER methods based on deep learning,the Bi-LSTM-CRF model has become one of the most popular NER methods because of its excellent ability of extracting contextual features.This dissertation considers that the Bi-LSTM-CRF still has some deficiencies,such as inabilitiy to distinguish polysemous words,distraction of attention and lack of features extracting from the local space of sequence.Therefore,this dissertation takes the Bi-LSTM-CRF model as the research basis,and uses TL-based(transfer learning)and EL-based(ensemble learning)methods to improve it from many aspects.At the same time,the recognition effect of People's Daily corpus is used as the evaluation standard to berify the feasibility of the improvement made in this dissertation.The main work of this dissertation is as follows:(1)After reviewing the development of NER and summarizing its status quo,the current shortcomings and defects of Bi-LSTM-CRF model are analyzed from various aspects.(2)Based on the idea of transfer learning,our research combines theBi-LSTM-CRF model with the pre-training model RoBERTa(Robustly Optimized Bi-direcional Encoder Representations from Transformers pre-training Approach),and proposes the improved model Ro BERT-Bi LSTM-CRF.The improved model solves the polysemous problem of Bi-LSTM-CRF by generating dynamic word vectors,and solves the problem of the distraction by using self-attention mechanism.Experiments on the People Daily corpus shows that the F-sorce,the evaluation standard,of improved model reached 94.92%,which was 5.28% higher than Bi-LSTM-CRF.(3)To slove the problem that the RoBERTa-Bi LSTM-CRF still lack the local spatial features,this dissertation studies the Iterated Dilated Convolutional Neural Network(IDCNN)which can extract local spatial features and can face to the problem of squence's long-term dependencies.Then,this dissertation integrates an IDCNN-based model with the RoBERTa-Bi LSTM-CRF model with the idea of ensemble learning,this work gets a model with better performance.Experiment on the EL-based model shows that the F-sorce of EL-based model is 6.18% higher than that of the Bi-LSTM-CRF model,which verifies the effectiveness of the El-based model proposed in this dissertation.