Research Of Chinese Named Entity Recognition Based On Attention Mechanism And Multi-source Embedding

Name Entity Recognition(NER)is a challenging natural language processing task,which is intended to identify entities that have practical meaning in text and classify the entities.To solve the problem that neural network based-methods are difficult to effectively cover text semantic features and capture key information,on the basis of BiLSTM-CRF model,this paper proposed a Chinese named entity model based on attention mechanism and multi-source embedding to better complete the task of named entity recognition.The specific research contents are as follows:(1)A Chinese named entity recognition model based on multi-source embedding(MSE-BiLSTM-CRF)is proposed to enrich the semantic information contained in the input of the model,and solve the problem of word boundary ambiguity caused by using single word vector.This paper takes BiLSTM-CRF model as the benchmark model,and introduces the multi-source embedding.Firstly,the trained character vector and word vector are spliced as the input of the model.Then,BiLSTM network coding is utilized to capture the semantic features of vectors.Finally,CRF is utilized to learn the direct relationship between adjacent labels and output the optimal annotation sequence.The experimental results on "People's Daily" news corpus show that MSE-BilSTM-CRF model can improve the effect of named entity recognition task.(2)On the basis of MSE-BilSTM-CRF model,the attention mechanism is introduced to solve the problem that the key information in the context cannot be highlighted.Firstly,the attention mechanism is added to the training process of neural network to calculate the attention weight of output features in the encoding layer.Then,the outputs of BiLSTM model at different times are weighted and summed based on the attention weight,achieving the effect of highlighting the key information.As shown in the experimental results,compared with MSE-BiLSTM-CRF model,the accuracy,recall rate and F-value of MSE-BiLSTM-Attention-CRF model increased 1.82%,1.37%and 1.41%respectively.The experimental results show that the accuracy,recall rate and F1-value of MSEBiLSTM-attention-CRF model proposed in this paper obtained 90.10%,92.23%and 91.15%respectively on the 1998 NER Dataset of "People's Daily".In addition,the Fvalue of three types of entity recognition(person name,place name and organization name)reached 92.37%,90.81%and 89.27%respectively.Compared with existing models,the effect is remarkable.Table[14]Figure[23]Reference[63]...