Abstract: Aiming at the time-consuming and labor-intensive problem of obtaining large-scale multi-hop question and answer training dataset, we  proposed a multi-hop question generation model based on the contrastive learning idea. The model was divided into the generation phase and the contrastive learning scoring phase. In the generation phase, candidate multi-hop questions were generated by executing the inference graph. In the contrastive  learning scoring phase, candidate questions were scored and sorted through a candidate question scoring model without reference question based on the contrastive learning idea, and the best candidate question was selected. This model had to some extent narrowed the gap between unsupervised methods and manual annotation methods, effectively alleviating the problem of lacking a multi-hop question and answer dataset. The experimental results on HotpotQA dataset show that the multi-hop question generation model based on contrastive learning can effectively expand the training data and greatly reduce the cost of manually labeling data.

Key words: multi-hop question generation, machine reading comprehension, contrastive learning