The reference reconciliation problem consists in deciding whether different identifiers refer to the same data, i.e. correspond to the same real world entity. In this article we present a reference reconciliation approach which combines a logical method for reference reconciliation called L2R and a numerical one called N2R. This approach exploits the schema and data semantics, which is translated into a set of Horn FOL rules of reconciliation. These rules are used in L2R to infer exact decisions both of reconciliation and non-reconciliation. In the second method N2R, the semantics of the schema is translated in an informed similarity measure which is used by a numerical computation of the similarity of reference pairs. This similarity measure is expressed in a non linear equation system, which is solved by using an iterative method. The experiments of the methods made on two different domains, show good results for both recall and precision. They can be used separately or in combination. We have shown that their combination allows to improve runtime performance.