Déjà Vu? Decoding Repeated Reading from Eye Movements

Technical Summary: "Déjà Vu? Decoding Repeated Reading from Eye Movements"

1. Problem Statement

While reading is often viewed as an interaction with new linguistic material, humans frequently engage in repeated reading (rereading) to improve comprehension, recall, or enjoyment. Psycholinguistic research has long established that eye movements during a second reading differ from the first due to "reading facilitation" (e.g., faster reading speeds, fewer fixations, and fewer regressions).

However, previous research has largely focused on descriptive, aggregated analyses (averaging effects across all participants and texts). This paper addresses a gap in the literature: Is it possible to decode, at the level of an individual reader and a specific text, whether a reading instance is the first or a subsequent encounter?

2. Methodology

The authors utilize the OneStop Eye Movements dataset, which contains high-resolution eye-tracking recordings of English native speakers reading news articles. They operationalize the decoding problem through two distinct prediction tasks:

• Single Trial Task: Given one eye movement sample, predict if it is a first or repeated reading.
• Paired Trials Task: Given two eye movement samples from the same participant on the same text (in unknown order), determine which is the first and which is the second.

Modeling Approaches:
The researchers implemented three main modeling strategies:

1. Feature-Based Models (XGBoost): Uses 35 hand-crafted features, including standard psycholinguistic measures (fixation duration, skip rate, regression rate), word property coefficients (responsiveness to frequency, surprisal, and length), and Saccade Network Measures (graph-theoretic properties of the scanpath).
2. Neural Multimodal Models (RoBERTEye): An extension of the RoBERTa language model that incorporates eye movement information by projecting eye movement feature vectors into the model's embedding space. They tested both word-level and fixation-level representations.
3. Synthetic Data Augmentation: A novel strategy where the models are augmented with synthetic "first reading" scanpaths generated by E-Z Reader, a computational cognitive model. This provides a machine-generated reference for what a "typical" first reading should look like.

Evaluation Regimes:
To ensure robustness, the models were tested across three regimes:

• New Participant: Model has seen the text but not the person.
• New Item: Model has seen the person but not the text.
• New Participant & Item: Model has seen neither.

3. Key Contributions

• New Task Definition: Introduced the formal task of decoding repeated reading from individual eye movement trajectories.
• Synthetic Augmentation Strategy: Demonstrated a method to improve predictive accuracy by integrating machine-generated cognitive simulations (E-Z Reader) into the training pipeline.
• Analytic Tooling: Leveraged predictive modeling not just for classification, but as an analytical tool to characterize the role of memory and reading speed in eye movement patterns.

4. Results

• Single Trial Task: All models outperformed the "Reading Speed" baseline (which uses only total reading time), proving that eye movement patterns contain information beyond simple speed. The XGBoost model augmented with E-Z Reader scanpaths achieved the highest overall accuracy (~70.2%).
• Paired Trials Task: This was significantly easier and more accurate. The XGBoost model substantially outperformed both the reading speed baseline and the neural models, reaching an accuracy of 91.4%.
• Model Comparison: Interestingly, the feature-based (XGBoost) models generally outperformed the neural (RoBERTEye) models, suggesting that the specific psycholinguistic features identified by researchers are highly discriminative for this task.
• Fine-Grained Insights: Analysis showed that as a reading session progresses, the model's confidence in identifying "first readings" decreases, likely because reading speed naturally increases over time, mimicking the facilitation seen in repeated reading.

5. Significance

This work advances the field of eye-tracking in NLP by moving from descriptive statistics to predictive modeling.

• Scientific Significance: It provides empirical evidence that eye movements capture a rich, decodable signal of a reader's prior memory and cognitive state. It also validates the use of cognitive models (like E-Z Reader) for data augmentation.
• Practical Significance: The ability to automatically detect if a user has already read a text has profound implications for e-learning and adaptive educational technologies. For example, an intelligent tutoring system could detect when a student is rereading a passage and adjust the difficulty or provide support based on whether the student is struggling with new material or reinforcing known concepts.