In this paper, we present an approach to general Datalog programs in which positive and negative facts are allowed. The semantics of such programs is computed in two steps. At the first step, we interpret rules as classical implication, and treat negation as classical negation. Since inconsistencies can appear at this first step, we consider that the information provided by facts in the database is more sure than the information provided by rules. Thus, a fact can be considered as an exception to a rule. Derivation of facts is based on the consistent part of the database, which can be localized. When no more facts can be derived, the second step of the computation consists in a modified version of well-founded semantics. In particular, an unfounded fact is assumed to be negative only if this assumption implies no inconsistency.