"Chinese Instructional Behavior Dataset for Autonomy-Supportive Teaching Detection"

"This dataset is the first dedicated one for deep learning-based instructional behavior research, developed to support the automated detection of autonomy-supportive teaching behaviors. Its raw data comes from 23.4 hours of in-class audio recordings of teachers across 8 subjects (including computer science, mathematics, psychology, medicine, etc.) at a Chinese university. Through Automatic Audio Segmentation (AAS) and Automatic Speech Recognition (ASR) technologies, the audio data was converted into 5203 Chinese sentence-level texts.  \r\n\r\nSubsequently, two trained psychologists independently annotated these texts based on the \"autonomy-supportive vs. controlling\" instructional behavior classification framework from educational psychology. After excluding 299 samples with inconsistent annotations, 4904 valid texts were retained, with the Cohen\u2019s Kappa coefficient for annotation consistency reaching 0.818.  \r\n\r\nThe dataset covers 14 instructional behavior labels, categorized into 7 autonomy-supportive behaviors (e.g., \"allowing student talking,\" \"offering hints,\" \"communicating perspective-taking statements\"), 6 controlling behaviors (e.g., \"uttering directives\/commands,\" \"making should\/ought to statements,\" \"deadline statements\"), and 1 neutral behavior (\"teacher talk,\" referring to knowledge-focused speech without autonomy-supportive or controlling attributes). A notable characteristic of the dataset is its severe class imbalance, with the \"teacher talk\" class accounting for 4180 samples (the largest class), and the imbalance factor (ratio of the largest class to the smallest class) exceeding 200."