| With the development of the Internet medical industry,the intellectualization and informatization of medical treatment have become the development trend.The electronic medical record stores a wealth of patient diagnosis and treatment data,which is an important foundation for the realization of intelligent and digital medical information construction.However,most electronic medical records are stored in the form of medical text entered in natural language.These unstructured,redundant and highly complex text data are difficult to directly obtain standardized and valuable content,and it is also impossible to directly apply artificial intelligence algorithms for further mining and analysis.Therefore,the structuring of electronic medical records has become a hot research topic in the era of artificial intelligence,and named entity recognition is the first step to realize the structuring of electronic medical records.At present,with the development of artificial intelligence,more and more studies have shown that deep learning has achieved good classification results in named entity recognition applications,but the premise is that a large amount of labeled data is required as training data for deep learning models.In the medical field,due to the professionalism,irregularity,and complex structure of medical entities,the data labeling of medical entities requires medical experts to spend a lot of time and cost to complete.Therefore,medical named entity recognition based on a small amount of labeled data has become a current concern.In order to solve the above problems,this paper studies a medical named entity recognition method based on a small amount of annotated data,which mainly includes the following two parts:1)Proposed a method for identifying named entities in breast electronic medical records based on self-training learningIn the medical field,medical text contains a large number of nested entities and professional terms,which requires a high level of expertise for annotating experts,and the cost of manually annotating data is high.However,a large amount of unlabeled data is stored in electronic medical records,so we need to make full use of the feature information of unlabeled data.First,only a small amount of data is labeled to train the initial model,and the initial model is used to predict the unlabeled data to select and generate pseudo-labeled data,so as to make full use of the unlabeled data.Then a hybrid data enhancement method is proposed,which uses a combination of pseudo-label data and entity replacement for data enhancement.Finally,for the enhanced data,the attention discriminator is used as the discriminant model of the data to screen out the enhanced data that meets the model training for subsequent model retraining.The final experimental results show that the F1 value of our method is as high as 75.22%,which greatly improves the classification performance of the model entity.Compared with the existing self-training learning method,the F1 value is increased by 10.19%.2)Proposed an entity recognition method for breast electronic medical records based on multi-standard active learningThe method based on self-training realizes the recognition of medical named entities under a small amount of labeled data to a certain extent,but in the implementation process,the selection of unlabeled data adopts a random selection method,so the utilization rate of data is still not high enough.In actual clinical practice,there may be relatively few research objects（such as malignant tumors,rare diseases,etc.）themselves,and relatively few overall data collected.Therefore,a multi-standard active learning based on breast electronic medical record entity recognition method is proposed.The advantage of this method is that the total amount of data required under the same accuracy is less than that of the self-training learning model.Active learning methods can select unlabeled data with high-value information for semi-supervised learning,and select the most valuable examples through selection strategies to be labeled by experts to achieve less training data with the same accuracy.In this paper,the active learning selection strategy is determined based on three criteria: the amount of labeled data,the cost of sentence labeling,and the balance of data sampling.For entity recognition tasks,a more suitable uncertainty calculation method for medical texts and sentence tagging cost measurement rules are proposed.Incremental training is used to speed up the iterative training in the active learning process.Compared with the research method of breast electronic medical record entity recognition based on self-training,this method reduces the amount of data required by about 63.6% under the premise of achieving the same accuracy. |
