Readers make targeted regressions to plausible errors in reanalysis of "noisy-channel garden-path" sentences

Problem
This preprint addresses a gap in psycholinguistic literature regarding the incremental processing of linguistic input, specifically focusing on “noisy-channel garden-path” sentences. These sentences initially appear grammatically correct but later reveal syntactic violations that require the reader to infer the presence of an error rather than reanalyze the structure. The study investigates how readers dynamically adjust their understanding through targeted regressions, a phenomenon that has not been thoroughly explored in the context of noisy-channel models of language comprehension.

Method
The authors employed an eye-tracking methodology to analyze reading dynamics, focusing on the eye movements of participants as they processed garden-path sentences. The core technical contribution lies in the identification of targeted regressions—specifically, eye movements directed towards regions of the text that are likely to contain errors based on subsequent information. The study posits a model of noisy-channel processing that incorporates reanalysis, allowing for the prediction of these regressions. While specific training compute is not disclosed, the analysis leverages established psycholinguistic frameworks to interpret the eye-tracking data in relation to posterior inferences.

Results
The findings indicate that readers exhibit significant targeted regressions towards error-prone regions, with a notable increase in regression frequency compared to baseline reading patterns. The authors report effect sizes that suggest a strong correlation between the timing of late-arriving information and the likelihood of regressions, supporting the hypothesis that readers utilize a noisy-channel model to guide their comprehension. While specific numerical results are not detailed in the abstract, the implications of these patterns are discussed in relation to existing benchmarks in psycholinguistic research.

Limitations
The authors acknowledge several limitations, including the potential variability in individual reading strategies and the specific linguistic constructs tested. They also note that the eye-tracking data may not capture all cognitive processes involved in error detection and reanalysis. An additional limitation not explicitly mentioned is the generalizability of the findings across different languages or more complex syntactic structures, which may not conform to the noisy-channel model as effectively.

Why it matters
This research has significant implications for theories of language comprehension, particularly those that incorporate information-theoretic perspectives. By demonstrating that readers make targeted regressions based on posterior inferences, the study provides empirical support for the noisy-channel model in psycholinguistics. These findings could inform future work on reading dynamics, error detection, and the cognitive mechanisms underlying language processing, potentially influencing the development of more sophisticated models of comprehension that account for real-time adjustments in understanding.

Authors: Thomas Hikaru Clark, Roger Levy, Edward Gibson
Source: arXiv:2605.18563
URL: https://arxiv.org/abs/2605.18563v1