Abstract: Calculating the similarity between the texts of the four examinations of TCM, recommending existing medical records that are similar to the patient’s four examinations performance can effectively assist clinical decision- making and professional learning. The texts of the four examinations of TCM lack the standard of clinical terminology, and the flexibility and individuation of phrasing are common. In order to obtain the effective representation of the four examinations on the limited scale TCM corpus, combining the characteristics of the four examinations texts, the text is characterized from the two levels of considering the vocabulary sequence and weakening the vocabulary sequence. The sparse attention mechanism was used to focus on the key features and enhance the interpretability of the model. After that, the GBDT was introduced to capture a variety of distinctive four examinations feature combinations to accurately predict the similarity between the two. The proposed method was verified on the text data set of the four examinations texts of TCM. The mean square error and Pearsonr coefficient were 82.06 and 0.89 respectively. The experimental results show that this method can effectively improve the semantic representation of four examinations texts and eliminate the influence of some irrelevant features, and enhance the capture of the combined features of two four examinations texts, so as to improve the prediction results of the similarity between the four examinations texts.