Research Of Biomedical Named Entity Recognition Method Of BioBERT-based Hybrid Model

Biomedical named entity recognition is one of the most important information extraction tasks in natural language processing tasks.It aims to identify specific medical entities from biomedical texts through rules-based,deep learning and other methods.In most deep learning methods,due to the model has problems such as large amount of calculation and recognition effect.The main work of the paper is as follows:One is for the model calculation problem.Based on the research of BioBERT benchmark experiments,the paper analyzes at the aspect of mathematical logic and uses Inverse Square Root Linear Units(ISRLUs)optimization based on the BioBERT.Optimizing GELUs whose derivatives consist of exponents into ISRLUs whose derivatives consist of polynomials,which reduces the computational cost and optimizes the back propagation speed.Second,in view of the deficiency of lacking of obtaining contextual information of BioBERT,the paper combines BioBERT and bidirectional long short-term memory network and conditional random field and then propose a hybrid model using ISRLUs,and uses 4 medical public datasets to verify the entity of this method in recognition effect and computation cost.The third is to explore the local improvement method of GELUs without completely replacing the BioBERT activation function GELUs,replace the original tanh fitting and approximate calculation ()with two composite integration calculations,and avoid the problem of approximate calculation defects,which provides a reference for special functions fitting in neural network.