A DeBERTa-based model for legal text information extraction: leveraging feature enhancement and virtual adversarial training

This paper presents a legal text information extraction model predicated on DeBERTa. By integrating RoBERTa weight initialization, feature enhancement modules, and virtual adversarial training (VAT), it effectively addresses the deficiencies inherent in existing legal text entity and relation extraction technologies. After undergoing fine-tuning on the legal domain triple extraction task, the model can proficiently extract key triples. Experimental results unequivocally demonstrate that it outperforms extant methods in terms of precision, recall, and F1 score, showcasing robust generalization capabilities and enhanced robustness. Ablation experiments incontrovertibly validate the efficacy of each module, thereby laying a solid foundation for the application of legal text automated analysis.