A Novel Online Teaching Effect Evaluation Model Based on Visual Question Answering

Abstract

The paper proposes a novel visual question answering (VQA)-based online teaching effect evaluation model. Based on the text interaction between teacher and students, we give a guide-attention (GA) model to discover the directive clues. Combining the self-attention (SA) models, we reweight the vital feature to locate the critical information on the whiteboard and students’ faces and further recognize their content and facial expressions. Three branches of information are encoded into the feature vectors to be fed into a bidirectional GRU network. With the real labels of the students' answers annotated by two teachers and the predicted labels from the text and facial expression feedback, we train the chained network. Experiment reports a couple of competitive performance in the 2-class and 5-class tasks on the self-collected dataset, respectively.