Law Retrieval by Supervised Contrastive Learning Using the Hierarchical Structure of Law

Abstract

We study the information retrieval task to identify the relevant law articles for a problem statement on a legal issue in civil law. In recent years, the mainstream approach has been to calculate the similarity between the problem text and each article using pre-trained language models. However, such methods have a weakness in retrieving relevant articles that have low n-gram similarity scores with the probelm. In this study, we show that in such hard cases, the articles tend to be of the same class as articles with high n-gram similarity scores in the hierarchical structure of the Civil Code. From this observation, we hypothesize that by making articles of same class close to each other in the feature space, we could make it easier to retrieve the above mentioned hard articles. Our proposed method realizes this by supervised contrastive learning using the hierarchical structure of the Civil Code. Experimental results show that the proposed method achieves higher performance in retrieving the correct articles with low n-gram similarity to the problem.