‘I am too nervous to speak’: The impact of teacher questioning on L2 learners’ emotions in a synchronous online class based on AI emotion recognition

Emotional disengagement has attracted considerable attention in educational research, particularly research on online learning environments. Negative emotions can narrow learners’ attentional focus and suppress spoken output, making it difficult for teachers to gauge students’ actual oral abilities. This study investigates whether real-time recognition of learners’ facial emotions, combined with targeted questioning strategies, can foster more positive affect during synchronous online second language (L2) classes. Using a multitask convolutional neural network (MTCNN) for face detection and the VGGFace model for expression classification, we captured seven emotions from 61 Chinese undergraduates during one-on-one online interactions. Each teacher question was coded for type, mood, and subject pronoun. The results extend research on classroom discourse and positive psychology by providing objective, facial analytics evidence that subtle linguistic choices shape learners’ affect in real time. We propose practical guidelines – balancing display and referential prompts, favoring rapport-building moods, and employing inclusive pronouns – to help teachers harness emotion data without adding undue cognitive load. The implications for L2 course design and future multimodal emotion research are discussed.