Abstract: Many existing algorithms of attribute reduction begin at constructing decision table’s discernibility matrix,then convert non－empty objects’ conjunctive normal form into minimal disjunctive normal form.In order to correct the error of discernibility function method for attribute reduction based on Skowron’s discernibility matrix which turns out to be error for inconsistent decision table,this paper proposes an improved discernibility matrix and the computation method on it,which suits for consistent or inconsistent data and is with lower complexity,especial to inconsistent data,it will get sparser matrix,so it is a simple and efficient method for attribute reduction．