Abstract: This paper addresses the problems that occur when the independence assumption is made in multiple concept learning. Basically, these are non-learnability of concepts, decrease of predictive accuracy and increase of learning time. However, dropping such an assumption raises other issues concerning the ordering effect, the adoption of the notion of extensional coverage and the generation of mutually recursive rules. All these issues are discussed in the paper and a solution is proposed for the first of them. Specifically, statistical techniques are exploited in order to discover concept dependencies before starting the learning process. This multistrategy approach has been tested on a real-world problem, namely document understanding. For this experimentation, a first-order learning system, called INDUBI/CSL, has been employed. Experimental results confirm two hypotheses: firstly, that by taking into account concept dependencies it is possible to improve the classification accuracy, and secondly, that statistics can profitably be exploited to infer such dependencies.