Named Entity Recognition In Medical Texts Based On Recurrent Neural Network

With the continuous development of Chinese information technology,medical data has been experiencing explosive growth on the Internet,and how to e ciently and reasonably utilize these electronic medical data has become a key factor in integrating data resources and improving e ciency.Named Entity Recognition(NER)technology in medical texts can extract key information from electronic medical records,drug instructions and so on,which is of great importance in improving the level of medical informationization,promoting medical research,and improving the quality of medical services.However,there are still some main shortcomings in Chinese medical text named entity recognition.Firstly,the accuracy of the model is not high on complex Chinese medical text datasets.Secondly,high-performance models are typically more complex and less suitable for practical applications that require fast and e cient processing of large amounts of data.To address these issues,this paper conducts the following research.(1)Due to the problems of gradient vanishing and gradient explosion in traditional recurrent neural network models,they perform poorly in entity recognition tasks in the Chinese medical field.This article proposes a Bert-Bi GRU-CRF model based on recurrent neural networks.Compared to traditional models,its advantage lies in using Bert as a pre trained learning layer to better understand text context information and learn the potential semantics of medical texts;Then,Bi GRU is used as the feature extraction layer to improve the accuracy of entity recognition;Using CRF as the label decoding layer to improve the accuracy of entity annotation.Finally,this paper compares Bert-Bi GRU-CRF model with other models by adjusting various hyperparameter and ablation experiments,and obtains that the F1 value of this model on the Chinese medical data set has increased by at least0.2%,which is a significant advantage in the comparison experiments with various models.(2)In order to further optimize the performance of the model and solve the problems of long training time and high complexity,this article optimizes the model in two steps.The first step is to introduce the self attention mechanism in front of the CRF layer to improve the F1 value.The second step is to use the Mish activation function instead of the original GELU activation function to improve the consumption of model calculation.Through experiments on a Chinese medical text dataset,it was found that the complexity brought to the model by introducing both self attention mechanism and Mish function was offset by their interaction.The F1 value obtained not only increased by 1.25% compared to the original model,but also decreased by 2.9 seconds compared to the model that introduced self attention mechanism alone or switched to Mish function,thereby improving model performance and e ciency.