<p>Web of Science is one of the most important databases for accessing academic information. It has a complex classification system, and the rationality and accuracy of its classification system are critically important for the retrieval of academic resources and the promotion of research within disciplines. In response to this, we selected a dataset from the “Multi-disciplinary category” within the Web of Science database and explored a classification prediction method based on gradient saliency text feature extraction. In this paper, we use this method not only to re-predict the single-category labels in the Web of Science database, which improves the quality by increasing the accuracy of text categorization and annotation, but also experimentally demonstrate that this method can effectively predict multi-categories, which provides a basis for decision-making on document categorization. It is found that this method can not only effectively guide the classification of literature, but also measure the reasonableness of the classification of database categories, as well as determine the extent to which the papers published by journals match the actual categories of the journals by analyzing the papers included in the journals.</p>