Researches On End-to-end Generation Method For Question Generation

Question generation is a task which aims to create human-like questions by inputting text sequences.The rule-based method is not effective,on the one hand,because the rule-based method mainly converts the declarative sentence into a question sentence through a large number of artificially constructed rules.The generated question of this method is too simple.On the other hand,it is because the questions which rule-based methods generated cannot be as creative and diverse as human questions.In recent years,question generation based on sequence-to-sequence method has gradually become a new research hotspot in the field of natural language generation.At the same time,with the introduction of many pre-training models which is rich in semantic knowledge,pre-training and fine-tuning have achieved great success in problem generation.However,the question currently generated are still far away from the application.Therefore,it is of great practical significance to study the problem generation and improve the performance of the problem generation.Different from traditional rule-based methods,question generation methods based on sequence-to-sequence method and pre-trained language model are both end-to-end methods.This paper conducts relevant research on the end-to-end problem generation methods.We will conduct research from the following aspects:(1)Question generation based on dual encoder and gated self-attention mechanism.Existing problem generation methods are mainly based on a sequence-to-sequence framework,which encodes sentence semantics through an encoder,and then enters the encoded semantic information into the decoder to generate corresponding questions.In this paper,we propose a question generation method based on dual encoder and gated self-attention mechanism,focusing on the problem that question generation method based on sequence to sequence encoder cannot effectively extract the interactive information between the input sequence and the answer.(2)Question generation based on the maxout pointer and answer masking.In this paper,we propose a question generation method based on maxout pointer and answer masking,for the decoding process cannot generate high-quality problems in sequence-to-sequence framework.(3)Problem generation based on pre-trained language model encoder.When people ask questions,they must first understand the input information,and then use existing knowledge to reorganize the words and ask questions.Pre-trained language models are trained on large amounts of unlabeled data,compared with the sequence-to-sequence-based problem generation method,it has more a priori knowledge and can capture more information while understanding the meaning of the input sequence.Due to more knowledge,the generated question is more abundant.This paper introduces a pre-trained language model as an encoder to enhance the extraction of input sequence information and improve the effect of the question.(4)We designed and implemented an end-to-end problem generation system.Based on the research content of the first three chapters,an end-to-end problem generation system is designed and implemented for easy access and use,and the prediction results of the model are visualized.Through the actual use of the problem generation system,a more intuitive problem generation effect is obtained.