A systematic review of multimodal emotion and sentiment analysis based agentic AI frameworks in educational systems

Abstract

The use of agentic Artificial Intelligence (AI) techniques in educational settings has significantly transformed teaching methods, leading to the development of an intelligent system that can recognise and respond to students’ emotional states in real-time. Since multimodal Emotion Detection and Sentiment Analysis (ED&SA) techniques can extract learners’ emotional and sentiment information to interpret the learning process, these strategies can be extended to build agentic-AI systems. To carry out a systematic review, the proposed study adopts a PRISMA-based approach on scientific publications published between 2020 and 2025 from major databases, including Scopus, Web of Science, IEEE Xplore, ACM Digital Library, and ScienceDirect. As a result, 90 out of 150 relevant publications are considered after following the inclusion, exclusion, and quality assessment criteria. The findings from these selected articles reveal a substantial methodological evolution, beginning with Machine Learning (ML) and progressing to multimodal Deep Learning (DL), transformer-based, and agentic-AI frameworks. According to the study findings, unimodal systems reported the performance accuracy in the range of 60% to 70%, whereas multimodal frameworks produced accuracy between 75% and 82%. Modern transformer-based and agentic-AI systems further boost the performance accuracy by about 15% and thus outperform the conventional baselines. Further integration of Large Language Models (LLM) with multimodal learning transforms these systems into intelligent self-regulating agents that autonomously promote learning through adaptive frameworks, emotional guidance, personalised curriculum recommendations, followed by context-aware assessment in real-world educational settings. The overall research findings suggest a significant paradigm shift toward multimodal learning environments guided by Agentic AI, where perception, reasoning, and action are combined into a single framework for building an intelligent educational environment. The objectives of the future research should be to build a large-scale, culturally diverse, and domain-relevant education dataset to help develop real-time, ready-to-deploy explainable and privacy-conscious AI frameworks.