Neural Question Generation Using Question Type Guidance

Abstract

Question generation aims to generate questions from a context. The quality of generated questions is related to the context, the corresponding answer, and the question type. Most of previous works seldom pay attention to the question type. In this paper, we focus on how to exploit question type to guide the question generation task. The question type prediction task is introduced into our multi-task framework to discover the underlying relationship between the context-answer pair and the question. In additional, we also apply metric learning mechanism to improve the semantic relevance between the generated question and the context. Our model achieves a better performance than the baseline systems on SQuAD benchmark. Experimental results demonstrate that the guidance of question type is significant for question generation.