A Variational Reasoning Medical Dialogue Generation Approach Under Low-resource Corpus

With the improvement of production level,people’s life has been greatly improved,health has attracted more and more attention of the public,and people’s demand for timely and accurate medical diagnosis has become stronger and stronger.However,China has a huge population,and there are many problems in the medical environment at this stage.The most serious one is that there is a huge shortage of personnel in many medical posts,and due to regional development differences,the distribution of medical resources in our country is extremely uneven.Under the urgent medical needs of people,the current medical online diagnosis and treatment websites have sprung up like mushrooms after a rain,such as ChunyuDoctor,Haodaifu,Muzhiyisheng and so on.Many online consultations and medical diagnosis conversations are made public after anonymization and can be easily accessed on the Internet.At the same time,due to the efforts of many researchers and the development of entity and relation extraction technology,a large amount of commonsense knowledge and medical knowledge are very easy to obtain.Through the relationship between disease symptoms in the medical knowledge graph,people without a medical background can even roughly infer the disease based on the symptoms.Large-scale open medical knowledge is of great help in improving the diagnostic capabilities,diagnosis and treatment recommendations and prescribing capabilities of assistant diagnostic systems.At the same time,due to the large-scale data and the current development of deep neural networks,it is possible to construct an assistant diagnosis system by combining external medical knowledge in a data-driven manner.Medical dialogue generation aims to provide fluency and accurate responses to assist physicians to obtain diagnosis and treatment suggestions in an efficient manner.In medical dialogues two key characteristics are relevant for response generation:patient states(such as symptoms,medication)and physician actions(such as diagnosis,treatments).In medical scenarios large-scale human annotations are usually not available,due to the high costs and privacy requirements.Hence,current approaches to medical dialogue generation typically do not explicitly account for patient states and physician actions,and focus on implicit representation instead.We propose an end-to-end variational reasoning approach to medical dialogue generation.To be able to deal with a limited amount of labeled data,we introduce both patient state and physician action as latent variables with categorical priors for explicit patient state tracking and physician policy learning,respectively.This allows the model to use partial labeled corpus for semi-supervised learning.We define a variational Bayesian generative approach to approximate posterior distributions over patient states and physician actions.We use an efficient stochastic gradient variational Bayes estimator to optimize the derived evidence lower bound,where a 2-stage collapsed inference method is proposed to reduce the bias during model training.A physician policy network composed of an action-classifier and dual reasoning detectors,a context reasoning detector and a graph reasoning detector,is proposed for augmented reasoning ability.The use of explicit sequences of patient states and physician actions with multihop knowledge reasoning helps to provide more interpretable dialogue generation results.We conduct experiments on three datasets collected from medical platforms.Our experimental results show that the proposed method outperforms state-ofthe-art baselines in terms of objective and subjective evaluation metrics.Our experiments also indicate that our proposed few-shot reasoning method achieves a comparable performance as state-of-the-art fully supervised learning baselines for physician policy learning.