Dynamic Demonstrations Selection for Few-Shot Legal Argument Mining

The selection of demonstrations for few-shot learning plays a pivotal role in the performance of LLMs. In the legal domain, selecting these examples becomes especially critical, since they must be both informative and economical. We address legal argument mining by adopting dynamic selection strategies where a specific set of demonstrations is selected for each inference, and we compare them to static approaches where examples are chosen by experts or by the LLMs themselves. We experiment with 34 learning configurations over 3 different tasks, i.e., classification of argumentative components, type of premises, and argumentative schemes. We find that dynamic selection methods are better than static ones in all three tasks, suggesting that similarity is an important criterion in this domain.