Abstract: Aimed at the problem that the question and the answer in the current question and answer generation method do not completely match, a method for automatically generating question and answer pairs from Chinese text based on neural network is proposed. We used the method of named entity recognition and rules to extract keywords from the text to determine the topic of the problem, and used the neural network model of multi-view matching to generate the problem from the text, avoiding strong dependence on manual templates. We used the reading comprehension model according to the problem to generate answers with higher confidence. The results show that the quality of the generated questions is higher than that of the template-based method, and 80% of unmatched question and answer pairs can be filtered.