We investigate the mechanisms that allow people to successfully understand language given noise in the world and in their own perceptual inputs. We address two parts of this question. First, what knowledge do people use to make sense of language inputs that may have been corrupted? Second, how much of this knowledge is used while people are processing sentences? We conduct a sentence production experiment and an on-line reading experiment in order to answer these questions. Both experiments provide evidence that syntactic knowledge can drive top-down reinterpretations of word identities, as well as syntactic reanalyses that are incompatible with people's language input. In addition, Experiment 2 provides evidence that this knowledge is deployed on-line, as people process sentences.