Few-Shot Law Retrieval System for Supreme Court Cases

Processing legal documents and understanding the array of law sections has been a challenge for natural language processing. It is often difficult to determine what law section is applicable to a court case without consulting a lawyer, whose cost can be prohibitive for many. Our objective was to build a system that retrieves relevant law sections given the facts relevant to the court case. To this end, we collected the most complete dataset of Supreme Court Cases ever compiled in Thailand and built machine-learning and rule-based systems to extract facts from the verdict, particularly claim and defence. We also structured the document for the downstream NLP processes. In this paper, we describe a few-shot law retrieval system based on the claim as the input in order to handle the long-tailed distribution of law sections. Our system outperforms the standard supervised baseline and exhibits its generalisability to the law sections that are rarely seen or not contained in the training set