Constraints on Exchange Edits During Noisy-Channel Inference

Area of Science:

• Psycholinguistics
• Computational Linguistics
• Cognitive Science

Background:

• The noisy channel framework explains successful communication despite input errors.
• Comprehenders infer intended meaning using prior probabilities and noise corruption likelihood.
• Understanding conditions for considering word exchanges as noise is crucial for sentence processing models.

Purpose of the Study:

• To investigate if and when comprehenders consider word exchanges as a source of sentence corruption.
• To examine how syntactic category and sentence type influence the interpretation of implausible sentences.
• To compare interpretation strategies in yes-no questions versus explicit sentence correction tasks.

Main Methods:

• Five experiments processing German sentences (SO, OS, passive) with implausible meanings.
• Sentences were repairable by exchanging function words or nouns.
• Interpretation probed via yes-no questions (Experiments 1-4) and explicit correction (Experiment 5).

Main Results:

• Implausible SO and passive sentences yielded few nonliteral interpretations.
• Implausible OS sentences elicited many nonliteral interpretations, regardless of verb type crossed.
• Word exchanges were considered for function words (same category) but not nouns in yes-no questions, especially when implausible sentences were rare.
• Noun exchanges were frequent when explicitly asked to correct sentences.

Conclusions:

• Comprehenders utilize word exchanges to repair sentence meaning, but this strategy is constrained by task demands and word type.
• The interpretation of implausible sentences depends on whether the task involves implicit repair (yes-no questions) or explicit correction.
• Task-specific prior probabilities assigned to implausible sentences likely explain the differing results between yes-no questions and explicit corrections.