Most of Information Extraction (IE) systems are designed for extracting a restricted number of relations in a specific domain. Recent work about Web-scale knowledge extraction has changed this perspective by introducing large-scale IE systems. Such systems are open-domain and characterized by a large number of relations, which makes traditional approaches such as handcrafting rules or annotating corpora for training statistical classifiers difficult to apply in such context. In this article, we present an IE system based on a weakly supervised method for learning relation patterns. This method extracts without supervision occurrences of relations from a corpus and uses them as examples for learning relation patterns. We also present the results of the application of this system to the data of the 2010 Knowledge Base Population evaluation campaign.