Exploring Latin WordNet synset annotation with LLMs

This study explores the application of Large Language Models to populate synsets in the Latin WordNet, keeping a human-in-the-loop approach. We compare zero-shot, few-shot, and fine-tuning methods against an English baseline. Quantitative analysis reveals significant improvements from zero-shot to fine-tuned approaches, with the latter outperforming the baseline. Qualitative assessment indicates better performance with verbs and polysemous lemmas. While results are encouraging, human oversight remains crucial for accuracy. Future research could focus on improving performance across different parts of speech and degrees of polysemy, potentially incorporating etymological information or cross-linguistic data.