Tracing Engagement in Language Learning: A Multi-Source Approach to Performance and Predictors

Student engagement plays a critical role in promoting students' learning outcomes. Engagement unfolds over time and is shaped by multiple factors such as learner motivation, school context, and task design. Whereas engagement trajectories have been widely explored in higher education and lifelong learning contexts (e.g., MOOCs), such research on young learners, especially in primary and middle school, remains scarce. Moreover, little research is done on the relationship between such engagement trajectories and task and learner factors. The aim of this study is to (1) examine engagement both at the task level and across multiple learning sessions and (2) explore how engagement patterns are shaped by learners’ demographic, cognitive, and psychological characteristics, and how these patterns relate to task characteristics and learning outcomes. We will use students’ trace data from the learning system, tests, and questionnaires from 157 8th graders collected in a randomized controlled trial (RCT) during a 4-week classroom-based intervention. Students were randomly assigned to an experimental condition with interactive, GenAI-based vocabulary explanations embedded in authentic L2 texts or a control condition without the interactive vocabulary function. Learning gain was measured with pre- and post-tests. We will first perform latent class analysis (LCA) on behavioral features extracted from trace data to identify unique engagement states at the task level. Student engagement states will be then examined with task and leaner characteristics using mixed effect regression models. Sequential mining will be applied to construct and study students’ engagement sequences at the session level, which will be clustered using agglomerative hierarchical clustering to uncover the distinct engagement trajectories among students. Lastly, we will examine the relationship between identified engagement trajectory clusters, learner characteristics, and learning gains with mixed effect regression models.