Research On Chinese Medical Question Answering Based On Deep Neural Network

Question Answering(QA)has always been a key issue in artificial intelligence,and it is also the basic content of Turing test and the key technology to determine the level of machine intelligence.Medical problems have always been a hot issue of people's livelihood that attracts most people's attention.In recent years,the online medical question and answer community is very active.Some professional doctors answer patients' questions in the community,which enhances communication between doctors and patients,alleviates the suffering of patients,and accumulates a large amount of data of medical question answering.At this stage,however,most medical questions are answered by the human.The study of medical QA can effectively reduce the workload of doctors and alleviate the imbalance of medical resources.This thesis aims to study Chinese community question-answering techniques for the medical field.There are two technical difficulties with this issue.First,Chinese medical text data contains a great deal of medical terminology.How to process and represent these professional medical texts so that they can be better handled and understood by computers? Second,there are different ways in which questions and answers are expressed by patients and doctors respectively,they express in different ways.There are semantic differences between questions and answers,that is,"semantic gap".How to reduce the semantic gap and accurately match questions and answers?Aiming at the Difficulty One,this paper uses the representation of character embeddings,the character embedding is trained by Chinese character directly,without using Chinese word segmentation,which avoids the inaccuracy of Chinese word segmentation in medical text and reduces the negative cascade effect of this inaccuracy on subsequent semantic matching models.Character embeddings contain a certain amount of semantic information,which can significantly reduce the number of unknown words,reduce memory consumption and increase computing speed.In view of Difficulty Two,two different semantic matching models are proposed in the paper.(1)Multi-scale convolutional neural networks semantic matching model,which can extract local information from different scales.Chinese words usually consist of 2~4 characters.In terms of character embeddings,multi-scale model can effectively extract semantic information of texts from different granularity such as words,words and phrases,etc.which effectively improves the accuracy of question-answer matching.(2)Multi-level composite convolutional neural networks semantic matching model,which uses multi-level structure and multi-level model can extract higher-dimensional semantic features.Instead of stacking multiple models,the model uses the output of each layer as the semantic information representation of the text,which enriches the semantic information of the text.Its ability to understand the semantic information is better and the accuracy of semantic matching is also higher.In order to testify the validity of the model,the first public Chinese medical QA dataset(cMedQA)was constructed.Experiments show that the character embeddings respectively with the above two models are composed of two QA frame,which extract semantic information of texts from breadth and depth,and have strong pertinence to the Chinese medical QA,greatly improving the accuracy of QA matching.Compared with other shallow semantic models and matching-based or statistics-based models,models proposed in this paper has greatly improved the accuracy.The top-1 accuracy of this model is up to 18% higher than other classic deep neural network models.