Few-Shot Legal Text Segmentation via Rewiring Conditional Random Fields: A Preliminary Study

Alfio Ferrara, Sergio Picascia, Davide Riva

Abstract: Functional Text Segmentation is the task of partitioning a textual document in segments that play a certain function. In the legal domain, this is important to support downstream tasks, but it faces also challenges of segment discontinuity, few-shot scenario, and domain specificity. We propose an approach that, revisiting the underlying graph structure of a Conditional Random Field and relying on a combination of neural embeddings and engineered features, is capable of addressing these challenges. Evaluation on a dataset of Italian case law decisions yields promising results.