Enhancing Question Generation with Commonsense Knowledge

Question generation (QG) is to generate natural and grammatical questions that can be answered by a specific answer for given context. Previous sequence-to-sequence models suffer from problem asking high-quality requires commonsense knowledge as backgrounds, which in most cases not learned directly training data, resulting unsatisfactory deprived of knowledge. In this paper, we propose multi-task learning framework introduce into question process. We first retrieve relevant triples mature databases select with the conversion information source context question. Based on these informative triples, design two auxiliary tasks incorporate main QG model, where one task Concept Relation Classification other Tail Generation. Experimental results SQuAD show our proposed methods are able noticeably improve performance both automatic human evaluation metrics, demonstrating incorporating external help model human-like questions.