Abstract: Traditional folktales are vital carriers of cultural identity, yet preserving and analyzing them in the digital age poses challenges. In the Philippines, Biliran folk narratives embody deep cultural insights that demand linguistically and culturally sensitive approaches. This study explores deep learning techniques for analyzing thematic elements in these narratives while maintaining authenticity. Specifically, it examines how multilingual language models can be adapted to identify, interpret, and correlate themes within culturally specific texts. The research employed a progressive fine-tuning strategy in four stages, adapting mBERT, XLM-RoBERTa, and GPT models to Biliran narratives. The methodology included: (1) constructing a corpus of 232 paragraphs, (2) thematic annotation by three trained annotators (κ = 0.78), (3) designing model architectures that integrate cultural knowledge, and (4) evaluating performance through automated metrics and expert assessments. The ensemble approach outperformed baselines, achieving an F1-score of 0.86 and reducing perplexity by 59%. Thematic classification identified Cultural Values as the most dominant theme (f1 = 0.90). Statistical analysis revealed significant correlations, particularly between Cultural Values and Family Relationships (r = 0.73, p < 0.01). Cultural experts validated the models’ effectiveness, rating authenticity at 4.5/5.0 and thematic coherence at 4.6/5.0. The study contributes: (1) a methodological framework for culturally sensitive AI analysis, (2) empirical proof that progressive fine-tuning enhances model performance, (3) statistical insights into thematic relationships reflecting Biliran cultural systems, and (4) practical methods for adapting multilingual models to low-resource contexts. These findings advance cultural heritage preservation and AI applications in diverse settings.