Authors:
Brasser, Jan, jan.brasser@uzh.ch, University of Zürich
Tschirner, Chiara, tschirner@cl.uzh.ch, University of Zürich
Stegenwallner-Schütz, Maja, maja.stegenwallner-schuetz@hu-berlin.de, Humboldt-University of Berlin
Jäger, Lena A., jaeger@cl.uzh.ch, University of Zürich
Keywords: reading development, visual search, machine learning, interpretable AI,
Abstract:
Early identification of children at risk for reading difficulties is crucial to ensure timely intervention and individualized learning support, ideally from the very start of formal reading instruction (cf. [1]). However, traditional reading assessments rely on basic reading skills, and are challenged by variability in early word-level reading abilities and non-standardized reading instruction.
Our prior work demonstrated that gaze features recorded during a visual search task can successfully distinguish primary school students with low word-level reading comprehension abilities from those with average to high skills both at the time of recording and one year later by leveraging machine learning models (Neural Additive Models and Random Forests). Moreover, these interpretable models enabled the identification of key gaze features during visual search that correlate with early word-level reading abilities.
In this talk, we present how models trained on visual search eye movement data from beginning readers generalize to pre-readers. To obtain test data, we conducted a visual search experiment with nine children aged 6 at Swiss kindergarten institutions, using different types of symbols (letters and generic shapes). The task was identical to the one completed by the primary school children, whose data was used to train the models.
Our results show that model performance drops significantly when applied in this zero-shot setting, suggesting that visual search performance is influenced by the acquisition of early reading skills. This finding also suggests that different gaze features during visual search are predictive of reading development in pre-readers, compared to beginning readers. This finding aligns with a study that suggested that visual search reflects the efficiency and direction of the visual information processing [2]. From a practical perspective, our findings highlight the need for dedicated models and training data specifically for pre-readers in order to achieve reliable predictions of future reading development before the onset of reading instruction.
[1] Connor, C. M. D., Morrison, F. J., Fishman, B., Crowe, E. C., al Otaiba, S., & Schatschneider, C. (2013). A longitudinal cluster-randomized controlled study on the accumulating effects of individualized literacy instruction on students’ reading from First through third grade. Psychological Science, 24(8), 1408–1419. https://doi.org/10.1177/0956797612472204 [2] Ferre}, G., Mazzo}, S., & Brizzolara, D. Visual scanning and reading ability in normal and dyslexic children. Behavioural Neurology, 19(1-2):87–92, 2008.