Interpretable Multi-Label Rule Extraction from Financial Regulatory Texts Using Transformer Models: A Comparative Study on EURLEX
Year: 2025
Type: dissertation
Abstract: Increased complexities of the financial regulations require automated systems that can extract interpretable representations of formulated rules into legal texts. The following paper describes the use of transformer-based models such as BERT, RoBERTa, and FinBERT to carry out multi-label classification of regulatory documents. The consideration of the whole spectrum of the NLP pipeline, encompassing such aspects as data preprocessing, label binarization, feature engineering, model training, and evaluation, is carried out by applying EURLEX dataset of the annotated EU legal texts. A traditional rule-based classification technique is also formulated to be used as benchmarking of performance. Micro and macro F1 score, precision, recall and confusion matrices are utilized to conduct the model evaluation. Interpretability is used in the form of SHAP, LIME, and attention visualization methods, where the input of the model prediction can be seen and rule importance can be learned. The assessments characterize that the transformer models have bettered the predictive accuracy and labelling coverage than the rule-based basic one and can generate scalable, albeit not completely explicable, outputs. This study would be a contribution to the field of law NLP since it proves that transformer-based models are effective in classifying and understanding the structure of financial regulatory documents, and they hold practical potential to automate compliance and application of legal AI.
Source: NORMA
Author: Bhalilkar, Atharva Sanjay
Institution: National College of Ireland
Language: English
Subfield: Law
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7204323208