Folktales are scattered across historical newspapers and generally unlabelled, making their automatic identification a challenging text classification task. This thesis examines the automatic detection of folktales in the newspaper Kmetijske in rokodelske novice using large language models. The gold-standard dataset comprises 2,860 articles, 29 of which themselves narrate a folktale under the adopted criteria. We compare two commercial and six open-weight local models using three prompt variants. The best commercial model achieves a Matthews correlation coefficient (MCC) of 0.923, while the best unadapted open-weight local model achieves 0.810. Both attain perfect recall. Using knowledge distillation, low-rank adaptation (LoRA), and quantized low-rank adaptation (QLoRA), we transfer the commercial teacher model's judgments and rationales to two open-weight local models. All adapters outperform their corresponding base models, raising the best local MCC to 0.861. The perfect recall achieved by the best models on this sample suggests their potential for preliminary candidate selection, while false-positive predictions still require expert verification.
|
