Abstract: To meet the requirements of information fusion computing at character-level, a character-abstraction method based on generalization and reduction is presented based on the rough set theory. The method of generalizing the values of attribute is studied according to the concept of level tree. Then, based on the principles of minimum average relativity and maximum consistency factor, two kinds of reduction methods are designed owning to their special goals, i.e., reducing redundancy attributes and reducing relative mapping relationship. The designed method can overcome the existing shortcomings of narrow object range and bad effect of the methods within the rough set theory. The correctness and dependability of the method are verified by application example.