Abstract: The damage of infrastructure projects under the action of earthquakes may result in regional function paralysis, and the earthquake damage assessment of urban areas can provide important information for rescue and reconstruction work. To make quick and informed decisions and respond quickly to major disaster events, a rapid assessment of the spatial distribution and severity of building damage is required. To this end, this study attempts to use the natural language deep learning model BERT to classify and evaluate building damage based on textual descriptions of earthquake damage. GB/T 17742−1999 is used to pre-classify the damage status of individual buildings in the affected area. To verify the effectiveness of the proposed method, this study collects the data of more than 300 individual buildings after the 2016 M6.4 earthquake in Tainan and establishes a text word cloud database. The dataset consists of 343 buildings (121 labeled light damage, 128 labeled moderate damage, 94 labeled severe damage), of which 85% are used as training and validation datasets, and the remaining 15% are used for robustness testing. When the BERT model recognizes the GB/T 17742−1999 labels of the test set, the overall accuracy rate reaches 81%. This study is just an attempt to establish a new and efficient earthquake damage assessment method. This method can be used to classify and evaluate the earthquake damage text descriptions on media and social platforms, and provide information for emergency rescue decisions and save a certain amount of time for strategic deployment.