BNU-CMHE: A Chinese Multi-Modal Heuristic Education Dataset from Beijing Normal University

Heuristic teaching is an active teaching pattern which emphasizes elicitation. Corresponding high-quality sam-ples are primal for developing Artificial Intelligence technology for evaluating heuristic. Thus, we present BNU -CMHE, i.e., a Chinese Multi-modal Heuristic Education Dataset from Beijing Normal University. It includes 3 tasks, namely heuristic teaching action detection, recognitionand classification. It contains 16 videos with audio, about 756 minutes in total, recorded from real lectures. 671 teaching action clips are highlighted. A heuristic level is labeled to each video. Time span, action category, heuristic level and 11 education characteristic labels are given to each action clip. Different from other multi-modal action-related datasets, the textual information is much more important than videos and audios. The dataset not only introduces new challenges comparing with other action-related datasets, but also fills the gap in automated education heuristic evaluation. We also propose two text action classification baselines on linguistic modality, taking LSTM and ALBERT as skeletons respectively. The skeletons are well-tested and proved well-performed on other tasks, but lead to unsatisfactory results on our dataset, which just leaves much challenge. The dataset and codes are available on http://ics.bnu.edu.cnl?page_id=93.