The current rise of big language models has provided new ideas for researchers in natural language processing, search engines, life science research, and other fields, but big language models have disadvantages in terms of high resource consumption, slow inference speed, and difficulty in application in industrial scenarios, especially in vertical fields. To address this problem, a multi-scale CNN (convolutional neural network) and bi-directional LSTM (long short term memory) fusion model for Thangka question classification were proposed. This model fuses global and local features of the data to achieve the Thangka question classification task, with the global features reflecting the essential characteristics of the data and the local features focusing on the easily overlooked parts of the data, and the two were fused in a stitching manner to enrich the feature representation of the sentences. Experiments on the Thangka dataset and the THUCNews dataset show that the model is slightly better than the Bert model in terms of accuracy, 1/20th shorter in training time and 1/3rd shorter in inference time, and also shows good applicability and effectiveness on text classification tasks.