基于对比学习和预训练模型的临床诊断标准化OA北大核心CSTPCD

Aiming at the problems caused by the current phenomenon of large scale of standard diagnostic thesaurus,limited textual relevance and uncertain number of standard words in clinical diagnosis normalization task,a clinical diagnosis normalization method based on contrastive learning and pre-training model was proposed.First,the simple contrastive learning of sentence embeddings(SimCSE)model was trained with a combination of unsupervised and supervised methods,and the obtained model was used to recall the candidate standard words from the standard thesaurus.Then,the candidate word reordering and classification of term counts were carried out based on bidirectional encoder representations from transformer(BERT),and finally the results were obtained.Experimental results show that the recall rate of the combined unsupervised and supervised SimCSE method is 86.76%,which is higher than other methods,and the BERT model has significant improvement in several metrics compared with other models in the reordering and classification of term counts.The proposed method achieves an F1 value of 72.54%for prediction on the test dataset,which is a good performance in clinical diagnosis normalization.