Temporal network analysis of cognitive load and meta-cognition dynamics in human-AI conversations

Abstract This paper introduces a formal methodological and computational framework that integrates temporal network analysis, cognitive load (CL) theory, and metacognitive (MC) research to model how humans learn and adapt during artificial intelligence (AI)-mediated conversations. The proposed approach provides a mathematical formalization of conversational cognitive features and their temporal interactions, defining a unified architecture and a functional platform that can estimate CL and MC awareness directly from conversation transcripts, without the need for specialized sensors or intrusive measurements. Conversational dynamics are represented as a multiplex temporal cognitive network whose layers capture distinct relational dimensions (semantic, syntactic, emotional, cognitive, among others), and latent cognitive states are inferred using a probabilistic Hidden Markov (HM) modeling of multimodal conversational features, with parameters calibrated through Maximum Likelihood Estimation. Rather than focusing on empirical validation, this work establishes the theoretical foundations, feature modeling, and inference mechanisms, together with the data processing pipeline and visual analytics components required to investigate cognitive–metacognitive dynamics in dialog-based learning. The framework opens new perspectives for designing adaptive educational chatbots, cognitively aware conversational systems, and analytical tools to study learning and collaboration in human–AI interactions.