Previous research has found that comprehenders are willing to adopt non-literal interpretations of sentences whose literal reading is unlikely. Several studies found evidence that comprehenders decide whether a given utterance should be taken at face value in accordance with principles of Bayesian rationality, by weighing the prior probability of potential interpretations against the degree to which they are (in)consistent with the literal form of the utterance. While all of these results are consistent with string-edit noise models, many error processes are known to be sensitive to the underlying linguistic structure of the intended utterance. Here, we explore the case of exchange errors and provide experimental evidence that comprehenders' noise model is structure-sensitive. Our results add further support to the noisy-channel theory of language comprehension, extend the set of known noise operations to include positional exchanges, and show that comprehenders' noise models are well-adapted to structure-sensitive sources of signal corruption.