Abstract:

Modern adaptive learning environments require difficulty estimates that reflect how learners process text while reading. In order to implicitly identify parts of text associated with a higher overall cognitive load, we present a reading embedding model based on individual eye tracking data. To achieve this, we combine pre-trained German sentence and word embeddings with eye-tracking-based labels that summarise the overall reader's cognitive load and serve as a metric for sentence-level reading difficulty. For each sentence, we extract a sentence embedding from a pre-trained German Transformer encoder and aggregate the embeddings of all its tokens into a sentence-level word representation. We then compute a continuous difficulty score derived from fixation durations, fixation counts, pupil size, and blink count in the eye-tracking data. A small neural adapter then maps the concatenated sentence- and word-level representations into a lower-dimensional space and predicts the corresponding sentence-level difficulty score. Using a small proprietary eye‑tracking dataset, we show that the resulting embedding can distinguish sentence‑level reading effort for unseen items, enabling meaningful ranking of candidate rewrites as relatively easier or harder. Our model can be integrated into adaptive learning systems and large language model–based rewriting tools, where its difficulty estimates can inform the selection of alternative phrasings that facilitate the explanation of programming concepts with reduced cognitive load. We also outline the requirements for expanding the approach to larger datasets and evaluating it on a greater number of students.

Keywords:

Learning Analytics, Personalized Learning, Eye-tracking, Word embedding models.