In this study, Prasad, Boughanem, and Dkaki introduce MESc; “Multi-stage Encoder-based Supervised with-clustering”; a deep-learning hierarchical framework designed to improve the classification of large unstructured legal documents. The MESc framework leverages LLMs to process segmented documents and applies unsupervised clustering to approximate their structure.

• Focused on long and non-uniformly structured legal documents with tens of thousands of words.
• Compared standalone LLM performance with their integration into MESc.
• Analyzed intra-domain transfer learning capabilities for GPT-Neo and GPT-J models.
• Conducted experiments on legal documents from different regions, achieving significant performance gains over state-of-the-art methods.

The integration of hierarchical frameworks and LLMs holds promise for improving legal judgment prediction, and the MESc framework’s efficacy lays the groundwork for advanced analytics in the legal domain. Dive into the methodology and results in the full article.