Eye-Tracking (ET) data provide valuable insights into human language processing and have recently been employed across a wide range of fields, from psycholinguistics to NLP. In this work, we introduce CERVINO, the largest Italian eye-tracking-while-reading dataset to date, comprising data from more than 200 participants reading legal texts, recorded at 120Hz. Our analyses show that the dataset captures established psycholinguistic effects: reading times are influenced by word length, lexical frequency, and surprisal. We further evaluate the usefulness of the dataset in downstream tasks by fine-tuning BERT on ET measures. Results show improved alignment between model attention and human reading behavior while preserving downstream task performance. Moreover, combining our data with higher-frequency ET resources yields the strongest alignment results, suggesting that lower-frequency ET data provide useful cognitive signals. Overall, these findings indicate that lower-frequency ET data retain sufficient cognitive information to provide useful supervision for attention alignment and complement existing high-frequency resources.

Focus of this work is the prediction of reading times as the task is customarily dealt with in literature: that is, by collecting eye-tracking data that are averaged and employed to train learning models. We start by observing that systems trained on average values are ill-suited for the prediction of the reading times for specific subjects, as they fail to account for individual variability and accurately analyze the reading gestures of specific reader groups, or to target specific user needs. To overcome such limitation, that is to predict the reading times for a specific subject, we propose a novel approach based on creating an embedding to compactly describe her/his fixations. Embeddings are used to individuate readers that share same or similar reading behavior from a reference corpus. Models are then trained on values averaged over this subset of similar readers. Experimental results indicate that the proposed approach consistently outperforms its corresponding variants, in which predictions of reading times for specific readers are based on data from all subjects rather than from the most similar ones.