Introduction
Traditional theories of human sentence processing generally assume that readers are strictly incremental and "grammatically sane." Under this view, the human parser only considers syntactic structures that are globally possible given the preceding context. However, psycholinguistic research has identified a phenomenon known as "local coherence," where readers are slowed down by syntactic structures that are plausible in isolation but impossible in their actual sentence context. For example, in the sentence "The coach smiled at the player tossed a frisbee...", the phrase "the player tossed a frisbee" resembles a standard active sentence. Even though the preceding context makes this interpretation impossible, readers exhibit measurable processing difficulty.
This paper addresses two major gaps in the study of local coherences: their ecological validity in naturalistic text and the lack of scalable, principled computational models to explain them.
Methodology
The authors employ a dual approach involving empirical analysis of eye-tracking data and the development of a new information-theoretic model.
Empirical Analysis
To test for local coherence effects in naturalistic reading, the authors analyzed the Dundee Corpus, which contains eye-tracking data for participants reading newspaper articles. They constructed regression models to predict "first-pass" reading times using a "local coherence factor." This factor was defined as the probability of a word's part-of-speech (POS) tag given only the word itself (Experiment 1) or the word and its immediate predecessor (Experiment 2). The British National Corpus (BNC) was used to estimate these baseline "bottom-up" probabilities.
Computational Modeling
The proposed model treats sentence processing as a Bayesian belief update. It quantifies difficulty using Kullback-Leibler (K-L) divergence between two distributions:
- Prior (Bottom-up): The distribution of structures suggested by a local string of words, independent of global context.
- Posterior (Top-down): The distribution of structures for those same words once the entire preceding context is integrated.
The model posits that processing effort is proportional to the distance between these distributions—the more a reader must "correct" their initial bottom-up impression to fit the global grammar, the greater the difficulty.
Results
The empirical analysis confirmed that local coherence effects are reliably present in naturalistic reading, not just in controlled laboratory settings. In Experiment 1, increasing the local probability of a tag significantly reduced reading times. Experiment 2 showed that multi-word local patterns (e.g., a noun-verb sequence) had an even stronger impact on processing speed.
Quantitatively, doubling the probability of a tag in Experiment 1 reduced reading time by approximately 0.7 to 0.8 milliseconds. In Experiment 2, the effect was more pronounced, with doubling the probability leading to a reduction of over 4 milliseconds. The K-L divergence model successfully replicated the "super-additive" difficulty observed in previous laboratory studies, validating its ability to capture the underlying mechanism of local coherence.
Discussion
The findings suggest that the human brain is frequently "distracted" by local logic. Even when a global structure is known to be impossible, the parser cannot help but partially construct the most likely local interpretation. This challenges the notion of a perfect global parser and suggests that human sentence processing is influenced by competing local and global constraints.
By framing the problem as a Bayesian update, the authors provide a principled mathematical framework for predicting reading difficulty. This approach moves beyond the "black box" dynamical systems used in earlier research and offers a scalable way to analyze complex phrasal structures across diverse sentence types.
Limitations and Future Work
While the empirical tests used simplified POS-tag probabilities, the authors suggest that future work should apply the full K-L divergence model—incorporating complex phrasal trees—to naturalistic data. Additionally, exploring more sophisticated grammar types beyond Stochastic Context-Free Grammars (SCFG) may further improve the model's fit to human reading patterns.
Summary updated