In this paper, too few samples will lead to over-fitting, and in order to get consistent assumptions, the assumptions are too complicated, which is called over-fitting. Imagine that a learning algorithm produces an over-fitted classifier, which can 100% correctly classify the sample data (that is, if you give it the documents in the sample, it will never go wrong), but just to correctly classify the sample, its structure is so complicated and its rules are so strict that any document slightly different from the sample data is considered not to belong to this category.