Multidimensional Team Communication Modeling for Adaptive Team Training: A Hybrid Deep Learning and Graphical Modeling Framework.

Team communication modeling offers great potential for adaptive learning environments for team training. However, the complex dynamics of team communication pose significant challenges for team communication modeling. To address these challenges, we present a hybrid framework integrating deep learning and probabilistic graphical models that analyzes team communication utterances with respect to the intent of the utterance and the directional flow of communication within the team. The hybrid framework utilizes conditional random fields (CRFs) that use deep learning-based contextual, distributed language representations extracted from team members' utterances. An evaluation with communication data collected from six teams during a live training exercise indicate that linear-chain CRFs utilizing ELMo utterance embeddings (1) outperform both multi-task and single-task variants of stacked bidirectional long short-term memory networks using the same distributed representations of the utterances, (2) outperform a hybrid approach that uses non-contextual utterance representations for the dialogue classification tasks, and (3) demonstrate promising domaintransfer capabilities. The findings suggest that the hybrid multidimensional team communication analysis framework can accurately recognize speaker intent and model the directional flow of team communication to guide adaptivity in team training environments.