This study centers on the task of generating questions based on a source context and corresponding answer, with an emphasis on multi-hop question generation (MQG), which generates complicated questions that require reasoning over several documents. In MQG tasks, key phrases establish connections between paragraphs. However, previous works focus on designing advanced key phrase selection loss to guide model to generate multi-hop questions accordingly, while they do not note that selected key phrase information can be injected into the decoder to constrain generating process. Furthermore, they ignore that the generated questions should be semantically aligned with the selected key phrases. To mitigate this issue, we propose a dual perspective key phrase selection QG framework (DKSQG). Specifically, DKSQG includes two perspectives: Forward Perspective, which simultaneously selects key phrases and injects key phrase information into decoder to guide the generation process. Backward Perspective, which uses dual attention to integrate questions with context, selects key phrases from the joint representation to ensure semantic alignment between questions and these key phrases. Experimental results demonstrate that our method achieves significant performance in the key phrases coverage, with performance enhancement over strong baselines on the HotpotQA dataset.