Student-Generated Texts as Features for Predicting Learning from Video Lectures: An Initial Evaluation

The digital trails that students leave behind on e-learning environments have attracted considerable attention in the past decade. Typically, some of these traces involve the production of different kinds of texts. While students routinely produce a bulk of texts in online learning settings, the potential of such linguistic features has not been systematically explored. This paper introduces a novel approach that involves using student-generated texts for predicting performance after viewing short video lectures. Forty-two undergraduates viewed six video lectures and were asked to write short summaries for each one. Five combinations of features that were extracted from these summaries were used to train eight machine learning classifiers. The findings indicated that the raw text feature set achieved higher average classification accuracy in two video lectures, while the combined feature set whose dimensionality had been reduced resulted in higher classification accuracy in two other video lectures. The findings also indicated that the Gradient Boost, AdaBoost and Random Forest classifiers achieved high average performance in half of the video lectures. The study findings suggest that student-produced texts are a very promising source of features for predicting student performance when learning from short video lectures.