Evaluate and Visualize Legal Embeddings for Explanation Purpose

Deep learning models have recently demonstrated promising results on legal-related tasks. However, the performance of the models is not enough for us to have a clear understanding of their behavior and the reason for such performance. This paper is not another performance-oriented report. It focuses on explanation, which is a detailed analysis of different legal embeddings to provide insight into such methods. We propose two quantitative metrics to assess the quality of legal embeddings. In addition, the visualizations presented in this work allow us to understand and choose the most appropriate approach for a specific legal problem. The results of this paper help us to explain the state-of-the-art model in legal processing rather than accept their performance as black boxes.