Research And Implementation Of Question Matching In Question Answering System Based On Multi-task Learning

Insurance is a way of risk management,which is mainly applied to avoid the risk of economic losses.It is inevitable that the policyholder will have all kinds of questions to be answered in the early,middle and late stages of the insurance.At present,there are still a large number of insurance companies using call center to provide services.The model has the following two disadvantages:on the one hand,users need to wait for customer service answers,which is time consuming;on the other hand,the company needs to maintain a large customer service team,which increases the human resource expenditure.To tackle such issues,this thesis studies the key technologies of the insurance domain question answering system,and constructs the insurance domain similar question matching model(Insurance Question Answering,IQ A)based on multi task.Multi-task learning is firstly applied into the field of Chinese insurance,and the tasks of insurance entity recognition and question similarity discrimination are carried out simultaneously.The task of similar question discrimination is the main task,and the task of insurance entity recognition is the auxiliary task.Corresponding to the model proposed in this thesis,based on the insurance QA English data set,this thesis adds two sub task labels to form a Chinese insurance multi task learning data set.Through the experiment,in the auxiliary task,F1 can reach 96%,which is better than the contrast experiment.In the main task,the comparative experiment selected 9 models which are currently at the top of the list,including 5 traditional deep learning models and 4 pre training models,ACC and ROC are 90.44%and 93.88%,respectively.Compared with the pre-training model,it takes only 1/3 of the excution time.It demonstrate that the model based on multi task learning is feasible and effective.