A Personalized Learning Feedback System Driven by a Lexical Semantic Network

This study presents a personalized learning feedback system driven by a lexical semantic network, tailored for low-resource language education, with a focus on Kazakh-language instruction. The core mechanism is an error- triggered review process: when a learner answers an assessment item incorrectly, semantic textual similarity (STS) is used to retrieve and recommend concept-linked review materials, prioritizing targeted remediation over traditional scoring. The system operationalizes a tripartite Term–Lecture–Assessment (TLA) chain, scheduling review tasks based on forgetting-curve principles. To enable this, we translated the SemEval Semantic Textual Similarity Benchmark (STSb) into Kazakh (STSb-kk) and developed a Kazakh Natural Language Inference cor- pus (NLI-kk). Experiments with the LaBSE model, fine-tuned in two stages, achieved a Pearson correlation of 84.72% on STSb-kk, demonstrating robust performance in similarity-based retrieval. A WordNet-style lexical semantic resource for Kazakh was constructed via translation, with lighter manual validation than STSb-kk and NLI-kk. While STS is not suited for grading complex open-ended responses, it excels in search and recommen- dation tasks in low-resource settings. The system offers a scalable framework for personalized remediation, an adaptive scheduling algorithm, and applicability to disciplines such as mathematics and physics. Future enhance- ments will explore multimodal content integration and real-time difficulty adaptation.