This thesis explores the use of large language models (LLMs) for the automatic
recognition of narrative motifs in folktales. We begin by presenting
folkloristic theory on motifs, classification systems (ATU, Thompson), and
modern digital corpora. We outline key natural language processing concepts
and LLM architectures, with an emphasis on instruction-based learning. In
the experimental part, we fine-tune the Gemma 7B model on structured examples
linking stories and motifs. We evaluate several learning strategies
(full fine-tuning, LoRA, distillation using Gemini 2.5 Pro) and perform both
quantitative and qualitative evaluations. Our results show that LLMs can
effectively classify motifs, especially when the dataset is enriched with chainof-
thought explanations. This work contributes to the field of computational
folkloristics and opens pathways for further research.
|
