Research On Yes/no Question Answering Based On Transfer Learning

Question Answering using depth-learning has achieved great success,at the same time,in order to meet the diverse and refined needs of users,QA begin to be refined gradually,including entity,description and Yes/No.Yes/No QA is an important part of QA,and the research is relatively few.Therefore,more research on Yes/No QA is needed.Yes/No QA is to answer the Yes/No question that need to get a clear opinion polarity.At this stage,the main problems of Yes/No QA are it need more the reasoning ability of model,more data and less resources when developing a model.The current mainstream models mainly use simple depth model or keyword model for Yes/No QA tasks,This scheme can not capture the reasoning information of Yes/No QA,and it is difficult to obtain subtle semantic differences.At the same time,Yes/No QA has the characteristics of less real data and higher labeling cost compared with other QA fields.In the case of fewer samples,models in depth-learning face the shortcomings of difficult training and poor results.The model parameters with high accuracy are generally large,and there are problems such as resource consumption and time-consuming prediction.This paper is based on transfer learning to improve the effect of deep learning model on Yes/No QA.Multi-task learning and knowledge distillation are two kinds of transfer learning mthods,which transfer data and model respectively.the main research contents include:Propose Yes/No QA model based on Multi-task learning.We use the pre-training model to solve the lack of common knowledge in the initial model,the related dataset is constructed which is related to Yes/No Q.The model can use the related dataset for multi-task learning and migrate the reasoning ability from other related datasets.It solves the problem:Yes/No QA have a small amount of data,and the model of Yes/No QA needs more reasoning ability.Propose Yes/No QA model based on knowledge distillation.This paper constructs a large number datasets related to Yes/No QA,and design some lightweight models.Teacher model score the related datasets,this paper uses the scored data to train the lightweight student model,and improve the accuracy of the lightweight model.It solves the problem that deep learning model needs a lot of resources and is slow in real applicationDesign and develop an online Yes/No QA system.The above model is integrated into Yes/NO QA system,and multiple supporting models are used to improve the accuracy of Yes/No QA,which enhance the credibility of yes no question answering structure,and achieve better effect.