Entity Resolution is a crucial task to integrate data from different sources to identify records that represent the same entity. Entity resolution commonly employs supervised learning techniques based on training data of matching and non-matching pairs of records and their attribute similarities as represented by similarity vectors. To reduce the amount of manual labelling to generate suitable training data, we propose a novel active learning approach that does not require any prior knowledge about true matches and that is independent of the learning method used. Our approach successively identifies new training examples based on an informativeness measure for similarity vectors by considering their relationship to already classified vectors and the uncertainty in the similarity vector space covered by the current training set. Experiments on several data sets show that even for a small labelling effort our approach achieves comparable results to fully supervised approaches and it can outperform previous active learning approaches for entity resolution.