Neural Question Generation based on Seq2Seq

Abstract

Neural Question Generation is the use of deep neural networks to extract target answers from a given article or paragraph and generate questions based on the target answers. There is a problem in the previous NQG(Neural Question Generation) model, and the generated question does not explicitly connect with the context in the target answer, resulting in a large part of the generated question containing the target answer and the accuracy is not high. In this paper, a QG model based on seq2seq is used, which consists of encode and decoder, and adds the attention mechanism and copy mechanism. We use special tags to replace the target answer of the original paragraph, and use the paragraph and target answer as input to reduce the number of incorrect questions, including the correct answer. Through the partial copy mechanism based on character overlap, we can make the generation problem have higher overlap and relevance at the word level and the input document. Experiments show that our proposed model performs better than before.