Leveraging ensemble machine learning and multimodal video complexity for better prediction of video difficulty in second language

Abstract

Research into multimodal language learning has shown that video materials are engaging and effective in teaching and learning language skills. While previous studies have predominantly focused on how to make learning from video more effective, little research has been directed at estimating video content difficulty. This study, therefore, aims to examine the efficacy of developing predictive models using different configurations of ensemble machine learning approaches (averaging, bagging, boosting, and stacking) and video complexity features (acoustic, phonological, lexical, syntactic, and visual complexity features). A corpus of 640 videos that were subjectively rated for content difficulty by 322 English language learners was used to build and evaluate predictive models. The results demonstrated that ensemble machine learning models substantially outperformed baseline regression models in predicting video difficulty. The findings are discussed in terms of their applications in second language learning and assessment.

What makes video‐based academic lectures difficult for language learners to comprehend? The role of multimodal complexity

“…A total of 320 instructional videos were selected for this study from the corpus of second language video complexity (SLVC) (Alghamdi, 2021). The video lectures discussed different topics from a broad range of academic disciplines including humanities, social studies, education, and computer science (see Supporting Information S1: Appendix A.1 for details).…”

“…The video lectures discussed different topics from a broad range of academic disciplines including humanities, social studies, education, and computer science (see Supporting Information S1: Appendix A.1 for details). The difficulty of the instructional videos drawn from the SLVC corpus was rated by 279 B1‐level EFL learners using a rating scale developed in Alghamdi (2021) (see Supporting Information S1: Appendix B).…”