Hybrid XRXNet for multilingual legal text analysis: a resilient approach to classification and attack mitigation
Arockia Babi Reebha S., D. Saravanan · Stochastic Analysis and Applications · 2025
Background: As digital data in multiple languages increases, analyzing multilingual legal texts has become increasingly important. However, existing models often struggle with the complexity of legal language and cross-lingual understanding, making accurate classification and interpretation challenging. Objective: The study aims to develop a natural language processing multilingual legal data model (NMLDM) to address domain-specific legal terminology, adversarial and prompt injection attacks, data sparsity, and the interpretability and transparency of legal text classification. Methodology: The proposed NMLDM introduces XRXNet, a hybrid model combining XLM-R and XLNet transformer-based language models for robust multilingual legal text analysis. NMLDM integrates data collection, executes preprocessing with an improvised mBERT tokenizer, and performs feature extraction using legal-specific lexicons and linguistic features. XRXNet combines cross-lingual understanding and long-range dependency modeling to classify legal documents accurately while mitigating adversarial and prompt injection attacks. Results: The proposed framework shows exceptional performance across several metrics after evaluating NMLDM with existing models. The simulation results show that NMLDM attained an accuracy of 99.66%, highlighting its competence. Thus, the proposed model shows promise for analyzing multilingual legal documents.