Named Entity Recognition Of Online Medical Consulting Texts Based On Deep Learning

With the development and popularization of Internet technology,many patients choose to consult doctors about health-related issues through online medical websites.Using information extraction technology to automatically obtain important information from patients’ online medical consulting texts,and then search the answers from the professional medical knowledge base to automatically provide patients with professional medical answers,can be an effective way to reduce the workload of doctors.Among them,Named Entity Recognition(NER)is a significant step in information extraction task.Therefore,the research on NER of online medical consulting texts has important practical significance.This thesis is devoted to the research on NER of online medical consulting texts based on deep learning.Through the investigation of relevant research status,it is found that the following problems exist: 1)The research on NER of online medical consulting texts is still in a blank stage.2)In the case of entity labeling with character granularity,it is necessary to study how to effectively fuse the local features and global features of the character in a sentence in the model.3)At present,it is rare to use the pre-training features of Bidirectional language model and Mask language model at the same time to improve the performance on NER.4)At present,it is rare to combine the language model pre-training method and the multi-task learning method to improve the performance on NER.In order to solve these problems,the following works are performed in this thesis: 1)A high-quality labeled dataset about NER of online medical consulting texts is constructed to fill the blank of the current research in this field.2)According to the characteristics of online medical consulting texts,the MQNer model is designed.MQNer can better learn the local and global features of the character in a sentence.Experiments show that MQNer can achieve good performance on NER of online medical consulting texts.3)Based on MQNer,the LM_MQNer model is designed innovatively.LM_MQNer uses both the Bidirectional language model and the Mask language model to pre-train the online medical consulting unlabeled texts,and then combines the pre-training features of the two language models into the model,thus obtains the grammar and semantic information contained in the unlabeled texts from different perspectives.Experiments show that LM_MQNer effectively improves the performance on NER.4)Based on LM_MQNer,a multi-task model AMTL_LM_MQNer using adversarial mechanism is designed innovatively.AMTL_LM_MQNer not only use the language model pre-training method,but also uses the adversarial multi-task learning method that combines the task of NER of electronic medical record texts.Experiments show that AMTL_LM_MQNer further improves the performance on NER compared to LM_MQNer.