We have recently defined RetroWeb, an approach to reverse engineer the informative content of semi-structured web sites. This approach provides a description of the web site informative content at physical, logical and conceptual levels. At each level a meta-model is instantiated using a set of reverse engineering rules. This paper focuses on the naming process used to instantiate the meta-models. We introduce an algorithm that will improve the labeling itself by reducing the number of objects to name. This algorithm is based on the analysis of the dependencies describing the inclusion, exclusion and equality between sets of objects.