Enhancing Intelligent Tutoring Systems with Conversational Support, Affective Models, Proactive Assistance, and Knowledge Tracing

• Autores: Yuyan Wu
• Directores de la Tesis: Miguel Arevalillo Herráez (dir. tes.) , David Arnau Vera (dir. tes.)
• Lectura: En la Universitat de València ( España ) en 2025
• Idioma: inglés
• Número de páginas: 138
• Tribunal Calificador de la Tesis: Olga C. Santos (presid.) , Verónica Romero Gómez (secret.) , Pablo Alfredo Salvá García (voc.)
• Enlaces
  • Tesis en acceso abierto en:  TESEO  RODERIC 
• Resumen

  • Intelligent Tutoring Systems (ITSs) leverage artificial intelligence to provide personalized instruction at scale, but their effectiveness is often constrained by limited understanding of learner states, including emotions, support needs, and interaction patterns. To address these limitations, presents four complementary contributions to enhance ITS personalization: affective modeling, flexible dialogue, proactive assistance, and knowledge tracing, which collectively advance the capabilities of ITSs. These contributions are addressed separately but in a complementary manner and are explored using the Hypergraph-based Intelligent Tutoring System (HINTS) as a research platform, designed for arithmetic and algebra word problems.

    The first contribution focuses on preliminary work on affective modeling for ITSs, using physiological signals and facial expressions. For physiological signals, including electroencephalography (EEG), electrocardiography (ECG), and electromyography (EMG), Hidden Markov Models (HMMs) are used to classify emotional states along arousal and valence dimensions, with EEG showing notable improvements in recognition accuracy. For the visual modality, the POSTER model, originally proposed for discrete emotion detection, is adapted for facial action unit detection as POSTER-AU, outperforming existing methods on benchmark datasets such as DISFA and BP4D.

    The second contribution enhances dialogue system flexibility by improving natural language understanding (NLU) and entity–quantity mapping in HINTS. An ensemble of four deep learning models (Bi-model, DCA-Net, AGIF, E2EMG-CRF) combined with hand-crafted rules ensures semantic consistency between detected intents and entities. Validated on ATIS, SNIPS, NLU-Benchmark, and a custom AWPS dataset, it achieves high accuracy even in low-resource conditions. The entity–quantity mapping links extracted entities to the system’s solution encoding, enabling tracking of student problem-solving. Using an adapted TF–IDF algorithm, this method handles short inputs, strong context dependencies, and limited data, outperforming sentence embedding approaches.

    The third contribution addresses proactive hint provision for students. Many learners display suboptimal help-seeking behaviors, which can impede their learning. To mitigate this, Transformer4Help, a Transformer-based model, is proposed to predict a student’s need for assistance based on their problem-solving behavior. Evaluations on real-world HINTS data demonstrate the model’s effectiveness.

    The fourth contribution focuses on Knowledge Tracing (KT) for student modeling, tracking concept mastery over time from interaction histories. We propose ITAKT, an attention-based model incorporating a talking-heads mechanism for enhanced behavioral modeling and Item Response Theory (IRT) for interpretable educational parameters such as student ability and item difficulty. Evaluations on ASSISTments2012, ASSISTments2017, and Junyi datasets show that ITAKT outperforms state-of-the-art methods in both predictive accuracy and interpretability.

    These contributions reinforce the core capabilities of ITSs and advance research in intelligent education. They promote personalized learning by enabling more adaptive and student-centered instruction. Furthermore, they have the potential to enhance learner motivation, accessibility, and inclusivity, while providing academic impact by improving the flexibility, adaptability, and robustness of ITSs.