Research On Question Generation Method Based On Seq2seq Architecture

Whether for humans or machines,asking and answering questions are both important ways to test the ability of understanding the given document,and with the development of neural networks,the demand for high-quality question-answer pair data in many fields such as machine reading comprehension and dialogue systems is increasing.The purpose of the question generation task is to automatically generate a reasonable and relevant question for the given document and the target answer,so the question generation task has also become an important research field.The goal of this paper is to improve the performance of the question generation model and thus generate higher quality questions.This paper mainly focuses on the neural network model of Seq2 Seq architecture,and aims to four problems existed: exposure bias caused by the difference between the training and testing processes of the current Seq2 Seq architecture model,insufficient utilization of semantic information of the input documents,poor performance for long text input,and insufficient utilization of answer information.(1)Aims to the first two problems,this paper proposes a question generation method based on semantic enhancement and reinforcement learning,which enhances the document representation through semantic graph and makes full use of the semantic information rich in the input document;Add the reinforcement learning framework,use three special indicators for question generation tasks as the reward function,and fine-tune the model parameters during the training process to improve the performance of the model.The experimental results on Hot Pot QA data set which has high requirements for semantic understanding and reasoning ability of the model show that this method effectively improves the performance of BLEU＿4 and ROUGE-L.(2)In view of the latter two problems,this paper proposes a question generation method based on answer information fusion and joint training,adding content selection tasks to improve the performance of the model in the face of long text input;obtaining the context representation of answer information fusion through multi-stage attention interaction between answers and documents,and using the potential interaction between the two tasks through joint training,and joint training also alleviates the problem of exposure bias to a certain extent.Experiments on the two divided versions of SQu AD show that this method has been improved in indicators such as BLEU＿4,ROUGE-L,METEOR,which effectively improves the quality of generated questions.Finally,this paper applies the question generation method based on semantic enhancement and reinforcement learning and some of the baseline method to the question automatic generation system,which includes model selection,single question generation and batch question generation.