Context ameliorates but does not eliminate garden-pathing: Novel insights from latent-process modeling

Read the full article See related articles

Listed in

This article is not in any list yet, why not save it to one of your lists.
Log in to save this article

Abstract

A theoretically important claim in psycholinguistics is that in English, linguistic context largely eliminates garden-path effects in temporary complement clause/relative clause ambiguities (e.g., Altmann, Garnham, and Dennis, 1992, Avoiding the garden path: Eye movements in context. Journal of Memory and Language, 31 (5), 685-712). However, this strong claim, based on comparing mean differences between groups of conditions in a factorial design, reduces the conclusion to a overly simplistic binary one: either context affects garden-pathing or it doesn't. Mean differences are an aggregate of different, non-deterministic parsing steps that could be very different from trial to trial. As a result, there may not be a simple yes-no answer to whether context eliminates garden-pathing. Sentence comprehension has been argued to involve a cascade of probabilistically occurring latent processes (e.g., Paape and Vasishth, 2022); seen in this way, context-sensitivity may only be occurring probabilistically and in a graded manner. We present a hierarchical multinomial processing tree model that spells out several probabilistic latent processes. The model reveals new insights into the impact of context on garden-pathing, going beyond the simple question of whether context reduces processing difficulty on average. Modeling using new data from a bidirectional self-paced reading study with 319 participants shows that the context affects both first-pass attachment probability and reanalysis cost. Because a simpler account of context effects than the MPT model could be a model based on surprisal derived from large language models, we compare the MPT model's predictive fit to a model using surprisal computed from large language models. Results show that the MPT model, which assumes a mixture distribution of reading times, outperforms surprisal-based models at predicting garden-path effects.

Article activity feed