Recent progress in information extraction has enabled us to create large semantic knowledge bases with millions of RDF facts extracted from the Web. Nevertheless, the resulting knowledge bases are still incomplete or might contain inconsistencies, either because of the heuristic nature of the extraction process, or due to the varying reliability of the Web sources from which they were collected. One possible way of resolving both issues is to reinforce the knowledge base with deductive power by appending first-order logical inference rules, which help to describe and to further constrain the domain with which the ontology deals. In our work, we investigate learning these rules directly from the data using Inductive Logic Programming (ILP), a well known technique which lies in the intersection of machine learning and logic. Although powerful, ILP is inherently expensive as there is a combinatorial growth of the search space when constructing these rules, as the size of the background knowledge grows. In addition, the evaluation of each rule becomes more expensive, as the number of the training examples is rising. Apart from that, it is not always obvious how to automatically select positive and negative training examples needed for learning new rules over a noisy and incomplete knowledge base. Based on my Master thesis, this talk explores the issues involved when applying ILP in a noisy, incomplete and large knowledge base.