A Rational Model of Eye Movement Control in Reading

Klinton Bicknell, Roger LevyACL 2010 · OpenReview

Abstract: A number of results in the study of realtime sentence comprehension have been explained by computational models as resulting from the rational use of probabilistic linguistic information. Many times, these hypotheses have been tested in reading by linking predictions about relative word difficulty to word-aggregated eye tracking measures such as go-past time. In this paper, we extend these results by asking to what extent reading is well-modeled as rational behavior at a finer level of analysis, predicting not aggregate measures, but the duration and location of each fixation. We present a new rational model of eye movement control in reading, the central assumption of which is that eye movement decisions are made to obtain noisy visual information as the reader performs Bayesian inference on the identities of the words in the sentence. As a case study, we present two simulations demonstrating that the model gives a rational explanation for between-word regressions.

Introduction

Reading is not a simple linear scan of text; it is a process of active Bayesian inference where the brain constantly updates its beliefs about a sentence based on noisy visual input. A significant challenge in eye-movement research is explaining "between-word regressions"—the seemingly inefficient act of looking backward. While traditional models often treat these as motor errors or processing breakdowns, this paper argues that regressions are an optimal strategy for resolving uncertainty when new context makes previous interpretations less certain.

The Rational Model

The proposed model frames reading as a Bayesian inference problem, where the goal is to identify the sentence by gathering visual information. It consists of three primary components:

  1. Language Model: An unsmoothed bigram model (derived from the British National Corpus) provides the "prior" probability of word sequences.
  2. Visual Input Model: This component simulates realistic, noisy vision using an asymmetric Gaussian function to represent how visual acuity drops off from the center of fixation (fovea). The model receives noisy letter identities, meaning no word is seen with absolute certainty in a single glance.
  3. Control Policy: A decision-making heuristic based on two thresholds:
  4. Forward Threshold ($\alpha$): If confidence in the current character is high enough, the eyes move forward.
  5. Regression Threshold ($\beta$): If confidence in any previous character drops below this threshold—often because new context contradicts an earlier belief—the model initiates a regression.
The visual input model uses a Gaussian distribution to represent acuity, updating character probability vectors through Bayesian inference.

The system is implemented using Weighted Finite State Automata (wFSAs), allowing for efficient updates of the probability distribution over possible sentences as new visual samples are fixated. This allows the model to predict both the location and the duration of fixations.

Simulations and Results

The model was evaluated using sentences from the Schilling corpus, with the language model trained on the 500 most frequent words of the British National Corpus (BNC).

Pareto Optimality of Regressions

In the first simulation, the authors tested whether looking back is actually efficient. By comparing non-regressive policies ($\beta=0$) with regressive ones ($\beta > 0$), they found that regressive policies are Pareto optimal. For any non-regressive policy, there exists a regressive policy that is both faster (fewer timesteps) and more accurate.

Accuracy vs. Timesteps demonstrates that regressive policies (beta > 0) achieve higher accuracy in less time than non-regressive policies.

Optimization and the Delay Effect

In a second simulation using PEGASUS optimization, the authors searched for the best possible $\alpha$ and $\beta$ values across various speed-accuracy trade-offs. In every case, the optimal policy involved a non-zero $\beta$. Interestingly, the optimal $\beta$ was often higher than $\alpha$. This suggests that because of the time lag in eye movements, the model requires a high threshold to trigger a return, ensuring that the regression is still necessary by the time the eyes actually land.

Conclusion

By treating eye movements as a rational information-gathering process, this model provides a first-principles explanation for regressions. Rather than being errors, backward movements are a fundamental part of an efficient reading strategy, allowing the reader to maintain a coherent belief state about the entire sentence in the face of visual noise and shifting context. Future work may involve more complex language models and the use of reinforcement learning to discover even more sophisticated control policies.

Summary updated

Abstract

A number of results in the study of realtime sentence comprehension have been explained by computational models as resulting from the rational use of probabilistic linguistic information. Many times, these hypotheses have been tested in reading by linking predictions about relative word difficulty to word-aggregated eye tracking measures such as go-past time. In this paper, we extend these results by asking to what extent reading is well-modeled as rational behavior at a finer level of analysis, predicting not aggregate measures, but the duration and location of each fixation. We present a new rational model of eye movement control in reading, the central assumption of which is that eye movement decisions are made to obtain noisy visual information as the reader performs Bayesian inference on the identities of the words in the sentence. As a case study, we present two simulations demonstrating that the model gives a rational explanation for between-word regressions.
BibTeX
@inproceedings{bicknell2010rational,
  author = {Klinton Bicknell and Roger Levy},
  title = {{A Rational Model of Eye Movement Control in Reading}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2010},
  url = {https://openreview.net/forum?id=Hk-kKol_WB},
}

Loading related content...

Figure