Assessment is viewed as an important means to understand learners' performance in the learning process. A good assessment method is based on high-quality examination questions. However, generating high-quality examination questions manually by teachers is a time-consuming task, and it is not easy for students to obtain question banks. To solve this issue, this study proposes an automatic high-quality question generation system based on natural language processing and Topic Model. A two-stage test-question generation method (sentence selection and neural question generation) is proposed in this study. We apply multisource teaching materials to select declarative sentences, and then a neural question generation model called topic-embedding question generation (TE-QG) is employed to generate high-quality examination questions. This model is based on attention and the pointer-generator mechanism. The experimental results show that the sentence selection method can select sentences that meet the key points of the course, and the performance of the TE-QG model outperforms those of existing NQG models.