Purpose/Significance To solve the deficiencies in the semantic representation of medical texts, and to realize the clustering of the retrieval results of the PubMed database.Method/Process The paper proposes a method to construct a fusion matrix by using the Jaccard coefficient and TF-IDF. Similarity relations between phrases, documents, and the contents of phrases and documents are combined to construct a fusion matrix, and several clustering algorithms are trained to group a collection of documents from the PubMed database. Category annotations are created to describe the meaning of each category of clustered documents.Result/Conclusion Experimental results show that the fusion matrix-based clustering is superior in grouping the document sets, and the extracted high-frequency words in the category descriptions distinguish the meanings of the categories well, so the fusion matrix design is effective for clustering descriptions of academic texts.