Research On Chinese Named Entity Recognition Based On Deep Learning

The main task of Named Entity Recognition(NER)is to identify proper nouns such as names of people,places and institutions in the text.It has been widely used in information extraction,sentiment analysis,sentence analysis and other fields.With the rapid development of the Internet industry,a large amount of text will be generated by users' social interaction and entertainment on the Internet.In order to mine the effective information in the text,named entity recognition technology has become the focus of researchers in various fields.Most of the early researches on named entity recognition adopted the rule-based method.However,these rules are mainly aimed at a specific field and require a lot of manual design.It is difficult to apply the rules of a certain field to other fields.Deep learning methods can automatically acquire features from data,which has attracted extensive attention and research in NER field.At present,most Chinese named entity recognition models based on deep learning can be divided into two categories according to the basic units of language processing:character-based model and word-based model.However,both of these two methods have their own disadvantages: the character-based model cannot make full use of the potential word information,and the word-based model cannot eliminate the ambiguity of words caused by different word segmentation models.In addition,most of the NER models based on deep learning are implemented on the basis of long and short-term memory network(LSTM)and conditional random field(CRF).LSTM can obtain the context information of the sentence,but it cannot highlight the role of keywords in the text sequence.LSTM can effectively model serialized data,but the chain structure of LSTM cannot capture global information,which is susceptible to word ambiguity and lack of word boundary information.Based on the above analysis,this paper carries out research from the following two aspects,and the main work is as follows:(1)Aiming at the problem that the character-based model cannot make full use of the potential word information and the LSTM cannot highlight the role of keywords,this paper proposes a Chinese Named Entity Recognition Model(ATT-BILSTM-CRF)based on the attention mechanism.The potential word information is obtained by combining character information with word information,and the context feature information is obtained by using BILSTM.By introducing the attention mechanism to calculate the weight of words and combining with the training of neural network,the weight of keywords will gradually increase to highlight the role of keywords and effectively prevent the loss of important information.Conditional random field(CRF)was used to obtain the optimal labeling sequence.Experimental results on three CNER datasets show that the proposed model is more effective in naming entities.The efficiency of the proposed model is compared with the Ontonotes data set.Experiments show that the proposed model has shorter training time and faster convergence speed.(2)The chain structure of LSTM cannot capture global information and is susceptible to word ambiguity.In this paper,a Chinese Named Entity Recognition(GGNN)model based on gated graph neural network is proposed,in which Bert model is used to better represent each character,dictionary information is used to obtain the potential information between characters,and global nodes are introduced to capture the global information in sentences.GGNN model follows a neighborhood aggregation scheme,using graph neural network to construct the relationship between characters and words,through multiple iterations of aggregation,using words in the graph structure,all the matching words and the whole sentence multiple interactions to solve the problem of word ambiguity.Conditional random field(CRF)was used to obtain the optimal labeling sequence.Experimental results on three CNER datasets show that the proposed model is more effective in naming entities recognition.Comparative experiments on the Ontonotes dataset show that the proposed model can process long sentences effectively.