Research On Chinese Named Entity Recognition Based On Deep Learning

Named entity recognition is a fundamental technical tool for text processing,and deep learning is favored by researchers in the field of named entity recognition because of its powerful learning ability to mine deeper text features.Named entity recognition based on deep learning is usually treated as a sequence annotation task,while a single feature extractor is not capable of extracting features,and usually cannot consider the global and local information of the text comprehensively,and cannot meet the demand of global and local feature extraction at the same time.In addition,traditional input embedding tends to focus only on word-level feature representation,which is good for English text,but insufficient for Chinese text.In the Chinese named entity recognition task,although the word-based feature representation caters to the characteristics that Chinese utterances are word-based and do not have natural word separation,it ignores the phonetic and morphological features of Chinese characters and the global semantic information of the utterance,and there is a problem that it cannot represent the multiple phonetic meanings of Chinese characters and the multiple meanings of similar characters with similar meanings,which leads to the recognition of the NER model The effect of NER model is not satisfactory.To address the above mentioned problems,this paper proposes two Chinese named entity recognition models based on deep learning methods,and the main research works are as follows.(1)To address the problems of single feature type and inadequate semantic representation of traditional NER models,this paper proposes a Chinese named entity recognition model that incorporates multiple embedding representations(FMER-CNER).The model improves its input layer on the basis of the Bi LSTM-CRF model.In the input representation layer,word embeddings and sentence embeddings are generated by using the Baidu ERNIE pre-trained language model,and the representation of sentence-level semantic features is added while modeling word-level semantic features,and on top of that,additional phonetic and morphological features including Pinyin,Wubi and Sijiao codes are introduced to further enhance the semantic representation capability of the model by fusing multiple features.Finally,the Bi LSTM-CRF model is used to complete the task of extracting features and obtaining the optimal label sequence.(2)In order to fully fuse multiple features,a vector fusion layer is specially designed in the FMER-CNER model in this paper.It consists of a fully connected layer,a Bi LSTM network and a multi-headed attention mechanism.As a relatively independent tool-based module,its main function is not only to fuse two horizontally spliced matrices and keep the matrix dimensions the same as those of the single matrix before splicing,but also to extract the features of the two matrices through the Bi LSTM network and the multi-headed attention mechanism and obtain more critical features according to the importance of the information.The more critical features are further obtained according to the importance of the information.(3)In order to solve the problem of insufficient feature extraction capability of a single neural network,the Chinese named entity recognition model with enhanced feature extraction(EFE-CNER)is proposed in this paper.The model takes the fused embedding representation obtained from the FMER-CNER model as input,and improves the feature extraction link.The Text CNN and Bi LSTM network are parallelized and the features are jointly extracted,which not only fully utilize the ability of Bi LSTM network to model the global information of the context,but also take into account the advantages of Text CNN to extract local features,and the two complement each other,and then add a multi-headed attention mechanism to focus the key information for both of them respectively to enhance the feature extraction ability of the model in a multi-party joint manner.Finally,CRF decoding is used to obtain the optimal label sequence.(4)The two models proposed in this paper are experimentally validated on the MSRA Chinese dataset,and the experimental results show that both Chinese NER models proposed in this paper can effectively improve the results of Chinese named entity recognition,and the test results outperform the results of other comparative experiments,proving the effectiveness and superiority of the models.