The global interest in Korean culture has driven a rising interest in Korean language learning, particularly among Russian speakers. However, existing learning resources often fail to combine interactivity with reliability. While Large Language Models...
The global interest in Korean culture has driven a rising interest in Korean language learning, particularly among Russian speakers. However, existing learning resources often fail to combine interactivity with reliability. While Large Language Models (LLMs) offer conversational fluency, they are inherently prone to hallucinations, producing plausible but factually incorrect grammar explanations that are detrimental to self-directed learners. To address this critical reliability gap, this thesis presents the practical implementation, and evaluation of a Korean grammar learning chatbot powered by LLM and Retrieval-Augmented Generation (RAG). The proposed system fundamentally shifts the role of the LLM from an unconstrained knowledge source to a reasoning engine grounded in an expert-curated, dual-source knowledge base consisting of precise grammar definitions and context-rich lessons. The architecture employs a Multi-Agent System (MAS) where a Router Agent performs intent classification to delegate tasks to specialized worker agents: a Grammar Explaining Agent for complex conceptual queries and a Direct Grammar Search Agent for precise, dictionary-like lookups. The system is accessible via a Telegram bot interface, selected for its widespread adoption among the target Russian-speaking demographic. This interface interacts with an extensive, high-performance API that ensures architectural modularity, allowing for the seamless integration of alternative frontend choices in future iterations.
Empirical evaluation utilizing a semi-synthetically generated benchmark demonstrates the necessity of the RAG framework – the Grammar Explaining Agent achieved an answer accuracy of 90.8%, significantly outperforming standard zero-shot (77.5%) and few-shot (81.5%) baselines. The evaluation framework detailed in this research had demonstrated that the Hybrid Search strategy – enhanced by Hypothetical Document Embeddings (HyDE) and an LLM-based reranker – achieves superior precision for direct grammar lookups, whereas Cross-Encoder based reranking performs great in the context retrieval in RAG. A mixed-methods user study validated the system’s efficacy, with participants reporting high levels of trust and clarity. This research demonstrates that grounding AI in verifiable external knowledge is indispensable for educational tools, successfully bridging the gap between conversational AI’s potential and pedagogical requirements.