Deep Learning Information Extraction For Chinese Medicine

In the field of Chinese medical information extraction,due to the lack of entity labeling corpus,the traditional named entity recognition model is difficult to achieve better accuracy and F1 value.Therefore,transfer learning and multi-tasking learning are commonly used in this field to overcome the scarcity of annotated corpus.However,in the past work,when sharing task data and network layer,the named entity recognition model based on multitasking learning will appear some noise outside the task and ignore the interaction information outside the task.In the field of information extraction,the named entity recognition task and the relationship extraction task are usually regarded as two relatively independent sub-tasks.In the joint training of relation extraction and entity recognition,there are some problems such as error propagation and information redundancy.This paper analyzes the current research status at home and abroad as well as the existing shortcomings.From the perspective of neural network sharing and annotation strategy,a new multi-task model and entity relationship joint extraction model are proposed.The main work contents are as follows:1.According to Chinese medicine tagging corpus is scarce,and ignore the problems existing multitasking model,put forward a kind of cross multitasking named entity recognition model of share structure,used to get target independent tasks outside of the interactive features,and use the language training model to enhance corpus,to improve the performance of the entity recognition.The model in this paper was tested on CCKS2017 and CCKS2018,and the F1 value reached 90.23 and 87.68 respectively,which proved the effectiveness of the model proposed in this paper.In addition,due to the problem of wrong segmentation of Chinese word segmentation tools in the field of Chinese medicine,this paper designed a named entity recognition model combining Chinese word segmentation tasks.This method can achieve a higher entity recognition rate on limited data sets,alleviating the problem of low entity recognition rate caused by word segmentation errors.2.In the task in the entity relationship extraction,in view of the traditional assembly line model of information redundancy,the problem of error propagation,labeling strategy to build a new decomposition and entity relationship extraction rules,Chinese stroke ELMO was introduced to the model of input layer model,and through the attention mechanism to relieve the labeling strategy categories imbalances.Experimental results of the model on Chinese diabetes dataset showed that F1 value of 72.17 was achieved,which achieved the best performance compared with other combined extraction models.3.Construction and implementation of Chinese medical question answering system.In this paper,the Chinese medical knowledge map was built by Neo4 j,and a question answering system was implemented on the Chinese medical knowledge map.The system integrates the functions of medical knowledge question-and-answer,entity identification query,medical knowledge retrieval,entity relation extraction query and knowledge graph visualization.