Research On Medical Text-oriented Machine Reading Comprehension

Machine reading comprehension technology is one of the frontier research tasks in the field of natural language processing,aiming to reason the answer to the question from the texts.Under the background of "Internet+",the combination of artificial intelligence and healthcare is the trend of the times.Studying machine reading comprehension technology in the medical field can help solve the problem of "difficult to see a doctor and expensive to see a doctor." However,there are currently few studies on machine reading comprehension in the medical field,and there is no large-scale machine reading comprehension dataset in the medical field.Therefore,researching machine reading comprehension technology in the medical field not only has great research value and practical use,but also has great challenges.This paper focuses on the topic of machine reading comprehension for medical texts.Firstly,it studies the single-text machine reading comprehension technology for medical texts.Then,in order to solve the problem of lack of data in the medical field,this paper proposes an automatic questioning technology for medical texts,which automatically generates a large-scale corpus in medical field.Finally,this paper focuses on the online medical search and answer scene,and studies the multi-text machine reading comprehension technology for medical texts.In the single-text machine reading comprehension experiment for medical texts,this paper first tried the BiDAF model,but it was limited by the lack of knowledge in the medical field,and the experimental results were not ideal.Therefore,this paper uses the medical knowledge that BERT learned in pre-training as external knowledge,and uses the BERT-based machine reading comprehension model which shows improvement according to the experimental results.The automatic questioning technology for medical texts includes two modules: sentence filtering and question generation.In the sentence filtering module,this paper proposes the Attention-based LSTM model and the BERT-based model.In the question generation module,this paper proposes a sequence-to-sequence model based on the attention mechanism,and introduces the copy mechanism to make the generated question more smooth,natural and real.Finally,this paper uses non-labeled medical texts to build a large number of machine reading comprehension training corpus in the medical field,and further improves the performance of the machine reading comprehension model on medical texts according to the experimental results.For the online medical search and answer scene,multi-text machine reading comprehension for medical texts is realized by combining single-text machine reading comprehension and candidate answer ranking.In this paper,based on the Pointwise method and the Pairwise method,several candidate answer ranking models are proposed,and various loss functions are tried in the Pairwise-based method.In the end,the experimental results of the multi-text machine reading comprehension task for medical texts were improved.