Fig. 7
From: Leveraging LLMs for legal terms extraction with limited annotated data
Impact of Normalized Levenshtein Distance Threshold (NLDT) on F1 score. An NLDT of O allows no variation between the annotations made by experts and the partial matches from LLMs. By permitting a sentence variation of approximately 60%, GPT-4 can achieve an F1 score of around 0.6