<![CDATA[Cultural heritage knowledge presents significant challenges for Question Answering (QA) systems due to their interpretive, context-dependent, and symbolically rich nature. While Transformer-based models have achieved strong performance in semantic representation, they remain prone to hallucination and contextual misalignment, particularly in culturally sensitive domains. This study proposes a Transformer-based cultural knowledge retrieval framework for domain-specific chatbots, combining a bi-encoder (MiniLM and MPNet) for efficient semantic retrieval and a cross-encoder (BERT-base) for fine-grained reranking. A curated dataset of 4,016 question–answer pairs in Indonesia is developed from cultural heritage sources and validated for contextual consistency. The proposed approach is evaluated using both quantitative and qualitative metrics, including accuracy, F1-score, Exact Match (EM), and semantic-based measures such as F1-BLEU, F1-EDIT, and F1-ANS. Experimental results show that while all models achieve high classification performance (accuracy up to 0.99), the BERT + MPNet configuration significantly outperforms others in answer quality metrics, indicating superior semantic fidelity. However, qualitative analysis reveals persistent issues of hallucination and contextual misalignment, highlighting the limitations of relying solely on statistical evaluation. These findings demonstrate that high numerical performance does not guarantee meaningful understanding in cultural domains. Therefore, this study emphasizes the need for hybrid evaluation frameworks and context-aware mechanisms to ensure epistemic fidelity. The proposed approach contributes to the development of more reliable and culturally grounded QA systems.]]>
Copyrights © 2026
