Deriving knowledge from information stored in unstructured documents is a major challenge. More specifically, binary relationships representing facts between entities can be extracted to populate semantic triple stores or large knowledge bases. The main constraint of all knowledge extraction approaches is to find a trade-off between quality and scalability. Thus, we propose in this paper SPIDER, a novel integrated system for extracting binary relationships at large scale. Through series of experiments, we show the benefit of our approach, which in general, outperforms existing systems both in terms of quality (precision and the number of discovered facts) and scalability.