We present a novel graph ranking model to extract a diverse set of answers for complex questions via random walks over a negative-edge graph. We assign a negative sign to edge weights in an answer graph to model the redundancy relation among the answer nodes. Negative edges can be thought of as the propagation of negative endorsements or disapprovals which is used to penalize factual redundancy. As the ranking proceeds, the initial score of the answer node, given by its relevancy to the specific question, will be adjusted according to a long-term negative endorsement from other answer nodes. We empirically evaluate the effectiveness of our method by conducting a comprehensive experiment on two distinct complex question answering data sets.