A Study Of Biomedical Event Information Recognition Method Incorporating External Knowledge

The use of natural language processing techniques to extract events such as drug treatment,disease diagnosis and other entities involved in the events from biomedical texts,as well as diseases,drugs and so on,is of great significance for academic research and various biomedical applications in the field of biomedicine.At present,researchers have conducted extensive research on biomedical event extraction techniques,but the following problems still exist in the identification of trigger words and relationship extraction in biomedical events:(1)The nested entities contained in biomedical events are difficult to identify,and the entity boundaries are difficult to determine;(2)There are a large number of abbreviations and domain-specific words in biomedical texts,resulting in the limited ability of traditional models to extract events and their relationships involving such abbreviations and proprietary words;(3)It is difficult to embed the semantic relations among entities involved in biomedical events.To address the above problems,this paper integrates a large external biomedical knowledge base and selfattention mechanism to investigate the task of biomedical information recognition extraction,and the main work is summarized as follows:(1)To address the problem that nested entity boundaries are difficult to identify,a biomedical event trigger word recognition model based on Sci BERT and self-attention mechanism is proposed.Firstly,text vectors,distance vectors and part-of speech vectors are fused in word-level encoding,and Sci BERT captures feature information of text words more extensively.This paper also uses multi-headed attention to weigh the importance of different features for entity recognition,multi-level capture of features to understand biomedical text,and fusion of richer semantic information to improve the performance of the biomedical trigger word recognition model.(2)To address the problem that abbreviations and specialized terms in biomedical texts are difficult to recognize,a pipelined biomedical event extraction model that integrates external knowledge and self-attention mechanisms is proposed.Firstly,we use Sci BERT training word vector for entity embedding construction,then we construct the candidate entity set according to the conditional distribution,then we use attention weights for word embedding weighting and local disambiguation of entities,and finally we get the final classification results through the output layer.The results show that external knowledge contributes to the performance improvement of biomedical event recognition models.(3)A joint biomedical information extraction model based on external knowledge and graph convolutional neural network is proposed to address the problem that biomedical semantic relations are difficult to embed.Due to the specificity of biomedical field,the existing model still faces the challenge of how to embed more semantic relations to assist the model for recognition and extraction.A graph convolutional neural network is used to construct local and external knowledge graphs respectively to automatically propagate and iterate the information carried by neighboring nodes.The feature information is propagated through the network topology to the node embedding by feature aggregation,and the interference of synonymous words is eliminated by using entity linking,and the experiments show that the joint task model improves the F1 value of event information recognition.