Research On Sequence Labeling Model Of Natural Language Processing Based On Deep Learning

Named entity recognition is a technology to extract entities with specific meanings from texts,and it is an important basic research task in the field of natural language processing.In Chinese news corpus,the performance of the existing named entity recognition model is reduced because of the polysemy in the text entity and the problem that different levels of features in long sentences are not fully integrated.With the development of artificial intelligence,using deep learning algorithm to improve the performance of named entity recognition model has become a hot research topic.Aiming at the problems that existing models can't solve polysemy and can't capture multi-layer features effectively,this thesis proposes a named entity recognition model based on attention-based multi-layer feature fusion.Firstly,the named entity recognition method based on context-related word embedding is studied.The method is based on Bi LSTM-CRF model,and proposes ELMo-Bi LSTMCRF model based on word embedding based on traditional named entity recognition model.ELMo dynamic word embedding is used instead of common word2 vec and glove static word embedding,and word embedding corresponding to each word is inferred according to context information to identify named entities more accurately.at the same time,the model has better performance in dividing named entity boundaries and identifying abbreviated entities.Secondly,on the basis of ELMo-Bi LSTM-CRF model,attention mechanism is combined with different models according to the meaning of each feature,word-level,word-level and sentence-level features are captured from three angles,and these features are fused,which makes full use of different multi-level features in Chinese text data,improves the quality of the model,and realizes named entity recognition based on multi-level feature fusion based on attention.In this thesis,the contrast experiment and ablation experiment are designed for the two improved models respectively.The experimental results on three Chinese news data sets prove that the two models have a certain improvement in the accuracy rate and recall rate compared with the current mainstream models in the task of named entity recognition.The ablation experiment proves that each module of the model can effectively improve the accuracy of model recognition.