| Question Generation(QG)refers to the automatic generation of questions from various inputs(text,database,images,etc.)by machines.As humans grow,they learn to ask rich,creative and revealing questions.The purpose of QG is to give machines the ability to ask pertinent questions,just like human beings.As a dual task of automatic question answering,QG is an important direction in Natural Language Processing(NLP),and has been widely used in the field of education,dialogue system and so on.However,Question Generation is a very challenging task.On the one hand,most of the existing work aims at generating the most possible question,but ignores that QG is essentially a "one-to-many" problem,that is,a given text can generate different questions,but there are few studies related to the diversity of QG,which has a great research value.On the other hand,using the correct interrogative words to ask questions is the basic requirement of generating a good question,but the mismatch between interrogative words and answer types is very common,seriously affects the quality of the questions.In order to solve the above problems,this paper takes the answeraware QG with paragraph-level contexts as the research object,and aims to improve the diversity of QG and the accuracy of interrogative word generation.The main work is as follows:(1)A paragraph-level question generation model based on variational attention is proposed.Inspired by the variational autoencoder,this paper makes a variational improvement on the attention mechanism from the sequence-to-sequence model,which changes the traditionally determined attention mechanism to be variational,different questions can be generated,thus increasing the diversity.At the same time,to effectively assess the quality of generated questions,a QA model based on Transformer is implemented,and its F1 value reaches 88.52.In this paper,the effectiveness of the proposed algorithm has been verified on SQuAD,a public question-answering dataset.Experiments show that the proposed model is superior to the benchmarks in both diversity and quality of questions.(2)A QG model combining interrogative word classifier is proposed.To solve the problem of inaccurate interrogative word generation,this paper models the interrogative word prediction task as an independent classification task,and designs an interrogative word classifier(IWC)based on BERT.Then,the classifier and the QG model are combined in a pipeline architecture.The classifier predicts the interrogative word that is provided to the downstream QG model.According to the context information,the QG model chooses whether to generate the interrogative word by using the copy mechanism,and finally the complete question is created.Experiments show that,compared with the single QG model,the QG model combined with IWC not only improves the accuracy of interrogative word generation,but also improves the overall quality of questions. |
