A dense video caption dataset of student classroom behaviors and a baseline model with boundary semantic awareness.