Research On Question Generation Based On Position Awareness Encoder And Question Types

With the development and application of deep learning and text generation,the field of machine reading has also made great progress.At present,most of machine reading comprehension datasets are in the form of question and answer.The expensive process of datasets seriously limits its size and domain,so the task of question generation has been a method to expanding data.It has also been widely concerned with academia.In this paper,all of question generation models was based on Seq2 Seq,the model takes paragraphs or sentences,with the answer which is cut off them,to generate rich semantics and strong diversity questions.most of the neural network methods use the Seq2 Seq framework by LSTM.But there are many problems,such as poor text representation,mismatching between generated question and answer,question including answer words,and single question generation form.To solve these problems,this paper studies on question generation based on encoder with fusing position awareness and question types,mainly at the machine reading dataset SQuAD develop the Seq2 Seq framework to algorithm research and the implementation of model.The work as following:(1)Focus on the problem at generated question has the words of answer.We proposed the model ‘hierarchical context representation with answer position awareness for question generation'.The Seq2 Seq framework consists of two component,one is encoder module with answer position awareness,the other is decoder module with coverage mechanism.We propose an answer position awaresnes encoder,it uses the word relative distance between paragraph and answer to differentiated encoding the text for words representation.By the way,for enriching the representation of text.The model using the paragraph from the original dataset SQuAD to expand the dataset at the task of question generation.Then model used encoder module to encoding the different level of text for getting the paragraph level and sentence level text representation.From that,model can get richer text information to generating questions.The two different level has been combined togenerating questions with the coverage mechanism on decoder module.The experiment show that the model can not only improve the quality of generated questions,but also effectively alleviated the probability of question words containing the answer.(2)In view of the single question generation from,we proposed the variation question generation model using the answer types.The model follows the idea of variational auto-encoder to learning using answer-type to generating questions.The data distribution of them optimized by KL_Loss under the different level text context,so that model can use the answer to generating question,also can change the answer type to generate diverse question.And for generated question cloud match the answer,the model add the module that using attention weight to differently adjusting the representation.Experiments show that the proposed model improves the task of question generation on dataset SQuAD,and enhancing the diversity of question.