Research On Named Entity Recognition Based On Deep Learning

With the development of the information age,the Internet not only brings convenience to people's life,but also produces a huge amount of data.Under the trend of data explosive growth,how to deal with massive non-structural data and extract effective information into the current most important issue.Named entity recognition technology can extract key entity information from massive text data.Therefore,named entity recognition task plays a vital role in this process,which has a high application value and research significance.In recent years,with the improvement of computer hardware conditions,GPU has developed rapidly.In the face of sufficient computing power,deep learning has been applied more and more widely.Without the limitation of computing power,deep learning is getting better and better.In the face of massive text data,deep learning technology can automatically extract effective feature information from it,avoiding manual feature extraction.In this paper,deep learning method is used to carry out named entity recognition task,and the main work and contributions are reflected in the following three aspects:Firstly,annotate the Chinese named entity identification data set.The experimental data in this article are derived from the Annotated Corpus of People's Daily(PFR)and the named entity recognition corpus of Microsoft Institute of Asian Research(MSRA).Both data sets are corpus that have been tagged with part of speech.For example,the names of persons,places and organizations shall be marked as nr,ns and nt.In this paper,the part of speech of the dataset is converted into BIO tags.BIO and BIOES are used to study the task of named entity recognition for two different data sets.Secondly,a named entity recognition model based on IDCNN-CRF is constructed.In the field of named entity recognition,convolution neural network is used to capture local information of text.In addition,with the deepening of the number of layers of the convolution neural network,the network parameters deepen exponentially.In order to meet this challenge,this paper constructs a named entity recognition model based on In order to meet this challenge,this paper constructs a named entity recognition model based on iterated dilated convolutions neural networks(IDCNN).Because IDCNN has no pooling layer,this model avoids data loss caused by up and down sampling during convolution.It increases the receptive field to extract broader global features and effectively solve the context dependency between long-distance sentences.In addition,this paper also uses the state transition matrix in CRF to learn the probability rules of the text output sequence and calculate the best output sequence label.This paper has done a lot of experiments on the IDCNN-CRF model,and determined a set of optimal experimental parameter configurations by adjusting different parameters.The F1 value obtained in two different data sets is 10.4% and 11.41% higher than that of the baseline model CRF,and 5.16% and 8.34% higher than that of the LSTM-CRF model,respectively.Compared with the Bi LSTM-CRF model,the effect is improved by 0.38%and 2.07%,and the training time is shortened by nearly 30%,which significantly improves the operation efficiency.Finally,a named entity recognition model based on BERT-Bi LSTM-CRF is constructed.Bi LSTM network can extract context information,but it can not represent polysemous words.In order to solve this problem,this paper introduces the word training vector of BERT on the basis of Bi LSTM-CRF.The BERT word vector with bidirectional Transformer coding effectively solves the representation problem of polysemous words through the Mask language model.BERT can extract the feature information at word level and sentence level,which greatly enhances the semantic representation ability of sentences.In this paper,BERT pre-training word vector is used to replace the traditional word vector representation,which provides strong semantic representation information for CRF.In the Chinese named entity recognition task of the "People's Daily" data set and the MSRA data set,compared with the baseline model CRF,the F1 value of the BERT-Bi LSTM-CRF model increased by 15.31% and 16.04%,and increased by 10.07% and 12.97% compared with the LSTM-CRF model.The F1 value of the BERT-Bi LSTM-CRF model is 5.29% and 6.7% higher than that of the Bi LSTM-CRF model,and 4.91% and 4.63% higher than the IDCNN-CRF model.