Weakly Supervision Based Slot Filling For Medical Diagnosis Dialogue System

In recent years,due to the growing demand for medical consultations,the medical consultation dialogue system has attracted more and more researchers’ interest.The medical consultation dialogue system is designed to collect personal information of patients through consultation and dialogue with patients,guide patients to state their own conditions,inquire about the missing symptom attributes in the patient’s statement,and finally provide diagnostic results based on all collected conditions,and Provide the corresponding treatment plan.The core module is the slot filling module in the semantic understanding module.The best performance on existing slot filling tasks is achieved on deep learning based models.The training of deep learning models requires a large amount of data.However,medical consultation dialogues are difficult to obtain and the cost of labeling is high,which limits the amount of labeling data.Limited labeled data also limits the performance of deep learning-based models.At the same time,in the online medical community,a large amount of unlabeled medical consultation dialogue data is generated by real doctors and patients.Unlabeled medical consultation dialogue data can help alleviate the problem of scarcity of expert labeled data.How to apply unlabeled medical consultation dialogue data to slot filling tasks is worth exploring.To improve the medical consultation slot filling task using unlabeled medical conversation data,This topic proposes the use of physician responses as weakly supervised information for patient condition statements.This method is mainly based on two observations:(1)Patients’ condition statements often use colloquial expressions,while doctors’ responses often use specialized words.Therefore,it is easier to obtain the medical entity words in the doctor’s reply by means of short-answer string matching;(2)There is a one-to-one correspondence between the patient’s condition statement and the doctor’s reply,that is,the doctor’s reply often contains the patient’s statement in the doctor’s reply.Paraphrasing of Medical Entity Words.Based on this,this topic uses medical entity words in doctor replies as weakly supervised labels for patient condition statements.With this approach,labels for patient condition statements can be constructed in an unsupervised setting,thereby constructing a large amount of weakly supervised data.This weakly supervised learning-based slot filling method can effectively improve the performance of the slot filling model by introducing a large amount of weakly supervised data.The research on the filling task of medical consultation dialogue slot based on weakly supervised learning technology is based on the weak supervision theory,which is the full use of unlabeled data,and has the advantages of simple method and easy practical application.Different from traditional supervised learning methods for slot filling tasks,the weakly supervised method proposed in this topic not only brings higher performance,but also opens up new research directions for medical dialogue slot filling tasks.In order to study this topic in depth,this topic first proposes a basic slot filling method based on weakly supervised learning.Then,in view of the problem of a large amount of noisy data in this method,this topic proposes slot filling methods based on weakly supervised data denoising and co-training methods.Finally,in order to adapt the method to more medical text semantic understanding tasks,this topic generalizes the method to phrase-level medical text natural language understanding tasks(medical entity normalization).Specifically,this topic has carried out research in the following four aspects:1.Wealy supervised learning based slot filling of medical dialogue systems.In order to solve the problem of data scarcity,this paper proposes a basic slot filling method based on weakly supervised learning.Specifically,this method obtains weakly supervised labels by matching the doctor’s reply with the medical knowledge graph,and the weakly supervised label and the patient’s condition statement form a weakly supervised data sample.The obtained weakly supervised data is used for slot filling model pre-training.The experimental results show that the method proposed in this paper achieves a significant performance improvement on the Chinese medical dialogue slot filling task.2.Weakly data denoising based slot filling of medical dialogue systems.Aiming at the problem that there is a lot of noise in weakly supervised data,this topic improves from the data perspective,and proposes a slot filling method based on weakly supervised data denoising.The method uses a self-learning mechanism and uses a small amount of expert labeled data to guide weakly supervised data denoising.Specifically,this paper uses a fusion mechanism to fuse the pseudo-labels and weakly supervised labels obtained by the self-learning method,aiming to eliminate labels with high noise probability.The weakly supervised data obtained after denoising is used for slot filling model pre-training.The experimental results show that the method proposed in this paper can effectively improve the performance of Chinese medical dialogue slot filling task.3.Weak-co-training based slot filling for medical dialogue systems.Aiming at the problem that there is a lot of noise in weakly supervised data,this topic improves from the perspective of model training method,and proposes a slot filling method based on weakly supervised data co-training.The collaborative training mechanism aims to learn data from multiple perspectives,and each perspective shares data with high confidence,so as to avoid the problems of unstable training and slow convergence caused by noise.The experimental results show that the method proposed in this paper greatly improves the experimental results without expert annotation data,and achieves a performance of85%.4.Weakly supervised based medical entity normalization.In order to apply the weakly supervised learning-based slot filling method to more medical text semantic understanding tasks,this topic generalizes the method to the medical entity normalization task.Aiming at the problem of data scarcity in the medical entity normalization task,this paper proposes a medical entity normalization method based on weakly supervised learning.The method extracts weakly supervised data for medical entity normalization task from weakly supervised data of medical consultation dialogues in an unsupervised manner.Experimental results show that our method achieves significant improvement on the Chinese medical entity normalization task.In general,this paper starts from the problem of slot filling in medical consultation dialogue based on weakly supervised learning,and proposes a slot filling method based on weakly supervised data denoising for the problem of a large amount of noise in weakly supervised data,and a slot filling method based on weakly supervised data Slot filling method for data co-training.Finally,the whole scheme is generalized to other tasks(medical entity normalization).