Controlling Reading Ease with Gaze-Guided Text Generation

TL;DR

This work tackles controllable reading ease in generated text by introducing gaze-guided text generation, which integrates a unidirectional gaze model predicting reading-time metrics into the decoding process of a language model. By applying gaze-weighted scores during beam search, the method can bias outputs toward shorter or longer reading times, with validation from an eye-tracking study on native and non-native English readers. Results show that reading times and perceived difficulty shift with the gaze weight, primarily driven by lexical features, and the approach remains modular and transfer-friendly since it does not require LM fine-tuning. The authors release a public gaze-informed reading dataset and provide a foundation for personalized educational content and accessible information, while noting limitations such as English-only data and focus on first-pass reading time.

Abstract

The way our eyes move while reading can tell us about the cognitive effort required to process the text. In the present study, we use this fact to generate texts with controllable reading ease. Our method employs a model that predicts human gaze patterns to steer language model outputs towards eliciting certain reading behaviors. We evaluate the approach in an eye-tracking experiment with native and non-native speakers of English. The results demonstrate that the method is effective at making the generated texts easier or harder to read, measured both in terms of reading times and perceived difficulty of the texts. A statistical analysis reveals that the changes in reading behavior are mostly due to features that affect lexical processing. Possible applications of our approach include text simplification for information accessibility and generation of personalized educational material for language learning.

Figures (7)

• Figure 1: Example application of gaze-guided text generation in a chatbot, and overview of the method. Our approach allows steering language model outputs to elicit certain reading behaviors -- for example, increased or decreased reading time. It integrates a gaze model predicting eye-tracking measures into the decoding process to modify the token probabilities generated by the language model.
• Figure 2: Gaze-guided text generation approach. (1) An off-the-shelf language model predicts a probability distribution (token scores) for the next token. (2) A fine-tuned gaze model predicts an eye-tracking measure (gaze score; in this case: FPRT, first-pass reading time) for each of the top $k$ candidates. (3) The gaze scores are multiplied with a user-defined gaze weight (in this case: $-2$) and added to the token scores. (4) The resulting total scores are used for decoding with beam search.
• Figure 3: Text characteristics across different gaze weights. Error bars show standard error of the mean from six generated texts. Points circled in red represent the texts shown to participants in the eye-tracking study.
• Figure 4: Length-normalized gaze and token scores predicted by the models, compared to first-pass reading times measured in the eye-tracking study. Token scores are token log probabilities predicted by the language model. Gaze scores are normalized first-pass reading times predicted by the gaze model. Error bars show standard error of the mean.
• Figure 5: Coefficients and 95% confidence intervals of linear mixed-effects models fitted to observed first-pass reading times. Upper panel: Model with gaze weight as predictor (compared to the reference with gaze weight 0). Lower panel: Model with word length and prevalence as additional baseline predictors. Both models include random intercepts for readers.