Modeling Working Memory Capacity in Sentence Processing with Lossy-Context Surprisal

Understanding how comprehenders interpret linguistic input in real time has long been a focus of sentence processing research. The recently-proposed Resource-Rational Lossy-context Surprisal (RR-LCS) model can be applied for making predictions at the level of individual readers: it predicts that individuals with low working memory capacity (WMC) have a lossy representation of the previous context and may therefore experience high surprisal at a long-distance dependency, while individuals with high WMC will experience a lower surprisal effect as this dependent to them may still be highly predictable from previous context. To date, these predictions have not been empirically tested. Here, we evaluate them using a controlled A-Maze task (Experiment 1) and an analysis of reading more naturalistic text based on the InDiCo corpus (Experiment 2). In Experiment 1, we investigated how WMC influences sentence processing across short, long, and very long dependency conditions. The results revealed that low-WMC participants initially showed a strong locality effect between the short and long conditions, which decreased over time due to adaptation. High-WMC participants demonstrated locality effects between both the short-long and long-very long conditions from the start, with these effects gradually diminishing as their memory allocation became more efficient. These findings align with the RR-LCS model's predictions, showing that high-WMC individuals maintained larger context sizes, such that they could more easily process long-distance dependencies, whereas low-WMC individuals, constrained by smaller context sizes, experienced pronounced locality effects. Experiment 2 confirmed that WMC modulates sensitivity to lossy-context surprisal (LCS), with LCS improving model fit particularly strongly for low-WMC participants. These findings underscore the significance of individual differences in WMC for processing complex syntactic structures. By combining experimental data with computational modeling, we provide a deeper understanding of how WMC shapes language comprehension.