In several areas of research such as knowledge management and natural language processing, domain-specific corpora are required for tasks such as terminology extraction and ontology learning. The presented investigations herein are based on the assumption that Wikipedia can be used for the purpose of corpus extraction. It presents the advantage of possessing a semantic layer, which should ease the extraction of domain-specific corpora. Yet, as the Wikipedia category graph is scale-free, it can not be used as it is for these purposes. In this paper, we propose a novel approach to graph clustering called BorderFlow, which we use and evaluate on the Wikipedia category graph. Additional possible applications of these results in the area of information retrieval are presented.