Evaluating knowledge-based approaches for legal text analysis: A benchmark study

Recent advances in transformer-based architectures and large-scale benchmarks have significantly improved Natural Language Processing performance. However, domain-specific areas such as law pose unique challenges due to specialized terminology, complex semantics, and limited annotated data. In this paper, we benchmark ASKE, a knowledge extraction approach designed to capture latent semantic structures in legal texts, against state-of-the-art legal-domain models using the LexGLUE datasets. We analyze its performance across diverse legal tasks through pseudo-precision and pseudo-recall metrics, highlighting its strengths, limitations, and adaptability. Based on the experimental results, we discuss potential application scenarios where ASKE can effectively support legal text analysis and downstream legal information processing.