Abstract: Existing multi-hop question generation models are usually focus on enhancing document representations. To further capture the semantic relationships between answers and contexts, this paper applies the pre-trained model RoBERTA encode documents and answers. To model the long distance dependencies, a semantic graph based on dependency analysis is then constructed to mine rich semantic relationships between texts. Finally, appropriate elements of the input sequence are selected as components of the problem via the maximum output pointer decoder. In addition, reinforcement learning is introduced to integrate syntactic metrics as a reward for augmenting the model training. Experiments on the HotpotQA dataset demonstrate that the proposed model achieves significant improvements according to BLEU1-BLEU4, METEOR and ROUGE-L compared with baseline models.