Bayesian Joint Modeling for Longitudinal Magnitude Data With Informative Dropout: An Application to Critical Care Data.

The way our eyes move while reading reveals important information about both the reader’s cognitive processes (Rayner, ) and the properties of the stimulus text (Rayner, ). Over the past decades, extensive research in cognitive psychology (Engbert, Longtin, & Kliegl, ; Reichle, Rayner, & Pollatsek, ) and psycholinguistics (Engelmann, Vasishth, Engbert, & Kliegl, ) has been dedicated to analyzing eye movements during reading and developing computational cognitive models to simulate human reading behavior. These models can be understood as computational implementations of theories about human reading behavior and its underlying cognitive mechanisms. The primary goal of this research line is to deepen our understanding of the cognitive mechanisms involved in reading and, more generally, human language processing. Most existing approaches remain heuristic, rule-based, or hybrid, as cognitive modeling prioritizes interpretability. Consequently, these models typically feature only a small number of free parameters to adhere to Ockham’s razor and to ensure falsifiability. While this enhances transparency, it often constrains their ability to capture complex patterns and adapt to variations across different populations, languages, stimulus layouts, lab setups, and reading tasks. Recent machine learning research has increasingly leveraged eye-tracking-while-reading for various technological applications. This includes enhancing natural language processing (NLP) models by incorporating eye movement features to enrich text features (Barrett, Bingel, Keller, & Søgaard, ; Hollenstein & Zhang, ; Mishra, Kanojia, Nagar, Dey, & Bhattacharyya, ) or to regularize neural attention mechanisms to make their inductive bias more human-like (Barrett, Bingel, Hollenstein, Rei, & Søgaard, ; Sood et al., ; Sood, Tannert, Müller, & Bulling, ). Eye movement data has also been employed to gain insights into the interpretability of neural language models (LMs), providing a means to better understand the differences between human and machine language processing and to evaluate the cognitive plausibility of LMs, which underpin state-of-the-art NLP systems (Beinborn & Hollenstein, ; Hollenstein, Pirovano, Zhang, Jäger, & Beinborn, ; Hollenstein, Gonzalez-Dios, Beinborn, & Jäger, ; Merkx & Frank, ; Sood, Tannert, Frassinelli, Bulling, & Vu, ). Additionally, eye movement data is utilized to infer a reader’s characteristics, such as cognitive conditions like ADHD (Deng et al., ) or dyslexia (Haller et al., ; Raatikainen et al., ), and linguistic skills, including reading comprehension capacity or whether the reader is a native speaker of the stimulus’ language (Ahn, Kelton, Balasubramanian, & Zelinsky, ; Berzak, Katz, & Levy, ; Reich et al., ). For a recent benchmark on predictive modeling from eye movements that includes most of these tasks, see (Shubi et al., ). Data scarcity has implications not only for training models but also for technological applications that rely on real-time gaze input for arbitrary stimuli at deployment. For example, the aforementioned approaches for enhancing LMs with gaze-derived features assume access to eye-movement data at runtime. In principle, such gaze-enhanced models could leverage human gaze patterns to identify relevant passages, for example, when generating a summary of an input text. However, the assumption that gaze data is available for arbitrary input texts is unrealistic for many real-world applications. For example, when a user wants a language model to summarize a given text, it cannot be assumed that eye-movement data for that specific text has been collected beforehand. To overcome this issue, researchers have started using simulated gaze data for enhancing LMs, resulting in advantages across various NLP tasks (Deng, Prasse, Reich, Scheffer, & Jäger, , ; Khurana, Kumar, Hollenstein, Kumar, & Krishnamurthy, ; Sood et al., ; Reich, Deng, Björnsdóttir, Jäger, & Hollenstein, ).