Constraints on Exchange Edits During Noisy-Channel Inference

According to the noisy channel framework of sentence processing, communication can succeed even when the input is corrupted because comprehenders rationally infer the speaker's intended meaning based on the prior probability of the literal interpretation and the probability that the input has been corrupted by noise. To test whether and under what conditions comprehenders consider word exchanges as a possible source of corruption, we ran five experiments on processing three types of simple German sentences: subject-before-object sentences (SO), object-before-subject sentences (OS), and passive sentences. Critical sentences had implausible meanings, but could be "repaired" by exchanging function words or by exchanging nouns. Experiments 1 through 4 presented sentences along with yes-no questions to probe interpretation. Implausible SO and passive sentences consistently elicited few nonliteral interpretations, whereas many nonliteral interpretations were given to implausible OS sentences. This was true regardless of whether word exchanges had to cross a main verb or an auxiliary, and it was more pronounced if the overall proportion of implausible sentences was low. We conclude that when answering yes-no questions, word exchanges are considered with function words of the same syntactic category, but not with nouns, and only when they result in a more likely syntactic structure. Experiment 5 showed that when explicitly asked to correct implausible sentences, comprehenders use noun exchanges frequently. We propose that the results for both yes-no questions and explicit corrections follow if the prior probability assigned to implausible sentences differs between tasks.