Experience feedback is of great significance to nuclear power plant construction. It can prevent the recurrence of problems and enhance the operational and engineering capabilities. However, currently, the nuclear power experience feedback data is complex, with judgment criteria difficult to quantify. Relying on a management model of subjective judgment, it poses great difficulties for screening tasks such as classification and deduplication. Moreover, there is a lack of effective mechanisms in terms of value mining and business driving. This paper is based on the pre-trained language model BERT. By introducing nuclear power industry corpus and using the Adaptive Hybrid Masking and Neighbor Product Reconstruction algorithms for secondary pre-training, the nuclear power professional model N-BERT is generated. Meanwhile, algorithms such as rank ordinal regression and fuzzy accuracy are designed. Through rule embedding and ordinal encoding, the discriminative ability of the model is enhanced, the hierarchical relationship is expressed, and the fuzzification of adjacent levels is achieved, significantly improving the screening accuracy of experience feedback. Ultimately, this paper clearly defines and quantifies the criteria for detection, deduplication, filtering, and grading. By eliminating subjective interference, solidifying the screening logic, it solves problems related to labels. The accuracy rate of some businesses exceeds 98%, reducing manual work and shortening the business response time. At the same time, with the help of N-BERT, a downstream application ecosystem is constructed, promoting the intelligent transformation of nuclear power business.