A robust method for collective disambiguation is presented, by harnessing context from knowledge bases and using a new form of coherence graph that significantly outperforms prior methods in terms of accuracy, with robust behavior across a variety of inputs.