Resarch On Question-answer System In Weak Supervision And Open Field Based On Knowledge Base

In recent years,a great deal of information has been generated on the Internet.It is important to get the information you need exactly and quickly from a large amount of information.Conventional information retrieval has been difficult to meet the needs of people's lives.Compared to the information retrieval system,the question answering system can provide people with accurate information quickly.However,using the method of strong supervision to train the question answering system requires a lot of annotation corpus,and the cost of annotating the corpus manually is very large.In this context,this paper presents a method of training question-answer systems based on the use of weak supervision methods in the knowledge base.It uses the Seq2Seq question generation model to generate question-answer corpora and then training the question-answer model with the generated corpus and a handful of annotated corpus.The main work of this paper includes the following two points:(1)This paper proposes using the Seq2Seq model to generate the question-answer corpus.Many existing research methods generate the question-answer corpus by means of templates and rules.The problems generated by the questionnaire are single and lack of the natural language diversity.However,the Seq2Seq model transforms the triples in the knowledge base into problems and makes full use of the knowledge information of the knowledge base.The generated problems have a better effect on several evaluation metrics than the methods based on the templates and rules.(2)This paper proposes to train the problem generation model and the question-answer model by joint training method.Compared with the method of first generating corpus and then training the model,this paper proposes a joint training language problem generation model and question-answer model.There is a certain deviation in the distribution of the corpus generated by the model and the annotated corpus.Training the corpus generation model and the question-answer model respectively can easily lead to the deviation of the question-answer model,and the suboptimal model can be obtained.By means of joint training,the mutual restraint between the two models can be strengthened and the performance of the model can be improved.Experiments on the SimpleQuestions dataset validate the effectiveness of joint training methods.