Research On Chinese Named Entity Recognition Based On Deep Neural Network

In today's social development process of mobile Internet information explosion,a large number of unstructured data has been created.It is a very important task to collect and sort these data,and use these data to extract and create value.As one of the principal tasks,named entity recognition extracts the characters of marked entities from a large number of text data and processes them structurally.As the fundamental work of language sequence processing,the accuracy of named entity recognition is of great significance to the implementation of related downstream tasks.Most of the early named entity recognition methods have poor recognition effect,heavy workload and poor generalization ability.At present,there are still many nodus for chinese named entity recognition.Nowdays,neural network has achieved outstanding achievements in many different fields,therefore,apply deep neural network to named entity recognition has great advantages.Based on the conventional chinese named entity recognition model,this thesis constructs the character vector representation through the Bert pre-training model,depends on its triple coding mode and pre training structurewhich the model can extract context feature information at character embedding level compared with the traditional model.In addition,it can solve the problem of disagreement of Chinese words.Extracting context feature information from text sequence using long short-term memory network.Obtain the label sequence through the conditional random field.The Bert-Bi LSTM-CRF model based on the deep neural network was established,by comparing the experimental results on three datasets,the Bert-Bi LSTM-CRF based on the Bert pre-training model can effectively improve the accuracy of Chinese named entity recognition,and its F1 value is greatly improved compared with the baseline Bi LSTM-CRF model.By replacing the long short-term memory with gated recurrent unit and introduce attention mechanism.Explore the performance effect of gated recurrent unit and attention mechanism through relevant experiments.Gated recurrent unit also has the advantage of long sequence modeling,and fewer parameters compared with long short-term memory,which can accelerate the convergence of the model on the premise of ensuring the effect of named entity recognition.Attention mechanism can locate the key character representation or feature information,which effectively improve the model recognition effect.Through improve the Bert-Bi LSTM-CRF model,a named entity recognition model based on BertBi GRU-Att-CRF structure is established.Compared with Bert-Bi LSTM-CRF,its F1 values on the three datasets are improved in varying degrees.