Abstract
Learning Management Systems (LMSs) increasingly need mechanisms that capture how learners respond during interaction, yet few platforms provide facial emotion recognition (FER) as a native, integrated capability. This paper presents the design of a real-time FER software module built directly into Moodle, together with its experimental validation. The module is a browser-based, client-side component that continuously estimates facial-expression confidence scores (happy, neutral, sad, angry, fearful, surprised, disgusted) while a learner interacts with a course, and links each reading to the participant, the interaction context, and a timestamp. To validate the module under realistic and varied conditions, 25 undergraduate students used it while interacting with four Moodle scenarios (E1, E3, E7, E9) built with contrasting user experience (UX) and gamification configurations, which here serve only as diverse interaction contexts for testing the module rather than as a design comparison in their own right. Two hypotheses were tested: that the module produces stable, non-degenerate dominant facial-expression readings, rather than noise, during interaction (H1), and that these readings differ significantly across interaction contexts (H2). The module produced a clearly interpretable dominant facial-expression reading in every context (record-level dominance between 51.28% and 73.55%, above the level expected under a simplified uniform-chance baseline), supporting H1. A Friedman test on a study-defined affective-valence indicator showed significant differences across the four contexts (χ2(3)=41.544, p<0.001, Kendall’s W=0.554), with five of six Bonferroni-adjusted Wilcoxon pairwise comparisons significant under the reported valence indicator (six of six under a robustness-check alternative), supporting H2, although the fixed scenario order means learning or familiarization effects cannot be excluded as a contributing factor. These results indicate that the module functions as intended and is sensitive to differences between interaction contexts; because the underlying pre-trained models were not independently benchmarked against the study population, this validates the module’s operation as a measurement instrument, not the accuracy of its facial-expression classifications against ground truth.
Funding Information
  • Universidad Politécnica Salesiana

This publication has 25 references indexed in Scilit: