More than words: Effects of grammaticality and lexical surprisal in self-paced reading

Language comprehension requires the integration of information from a wide variety of sources, including sensory input and memory. The present study contributes to a growing literature examining how probability and uncertainty shape language comprehension in close collaboration with grammatical knowledge, wherein the specifics of when and how these sources of information come together during sentence processing remain opaque. Here we asked how morphosyntax and lexical surprisal impact subject-verb agreement in Dutch: an online self-paced reading experiment tested whether lexical surprisal affects the use of grammatical information. Using a combination of model-comparison and permutation testing, we showed that reading times cannot be explained by a dichotomous interactive, nor by a purely additive account. While both factors must be modeled in order to best describe reading time data, the data provided evidence that lexical surprisal is leveraged more reliably when the constraints placed by the grammar are obeyed. We propose a novel "structure bottleneck", where the use of probabilistic cues is conditionalized on grammatical status of a subject-verb agreement relation. Viewed in the context of previous findings, the results are consistent with an account of language comprehension wherein grammatical and contextual probabilistic cues are weighted based on their reliability.