Handling numerical information is one of the most important research issues for practical applications of first-order learning systems. This paper is concerned with the problem of inducing first-order classification rules from both numeric and symbolic data. We propose a specialization operator that discretizes continuous data during the learning process. The heuristic function used to choose among different discretizations satisfies a property that can be profitably exploited to improve the efficiency of the specialization operator. The operator has been implemented and bested on the document understanding domain.