Abstract
Developing computational models of human behaviors for experts in many science fields has been at the forefront of several interdisciplinary research. In this work, we collaborate with researchers from National Academy for Educational Research (NAER) to develop an automatic scoring system for pre-service principals’ impromptu speech at the certification program. We propose a dense unit-level feature extraction and session-level encoding methods to characterize principals’ multimodal behavior. Moreover, we extend the framework by two direction. First, with inspiration from the psychological evidence in human’s decision-making mechanism, we assign confidence scores outputted from classifier as the predicted scores to all the speech. Secondly, as recent works on multi-task learning have been successfully utilized in many fields, we leverage other training tasks of the certification program to incorporate information about these new targets and achieve better performance of our scoring system. All the experiments demonstrate that our framework indeed has capability in handling high-level and subjective attributes.