Abstract: Typhoon disaster assessment is a multi-attribute decision-making problem. Aiming at the authenticity and timeliness of decision information sources after typhoon disasters, this paper proposes a typhoon disaster assessment method based on the masked language model as correction BERT (MacBERT) within a neutrosophic environment. Firstly, a fine-tuned MacBERT model is used to quantify the real-time online commentary information by indicator, and the single-valued neutrosophic numbers are used to express the classification results of typhoon disaster comments. Subsequently, technique for order preference by similarity to an ideal solution (TOPSIS) method is used to rank the impact severity of typhoon disaster across different regions, providing a prioritized sequence to support post-disaster emergency rescue operations. Finally, a case study using typhoon Hagupit is conducted. A detailed sensitivity analysis is performed to determine the optimal parameter setting for the classification model. Then, ranking results are compared with official statistical data, demonstrating the effectiveness of the proposed evaluation method. From the point of view on the data conversion efficiency compared with the exact number, it is proved that the single-valued neutrosophic numbers can retain the original data more effectively than the exact number when expressing the model classification results. The classification performance of MacBERT is also benchmarked against the ChatGPT model, confirming that the fine-tuned MacBERT model achieves better results in processing typhoon disaster commentary text, with an optimal F1-score reaching 0.983.