Automated Legal Entity Extraction with Legal-Bert and Ner

Abstract

The rapid growth of digital legal documents creates opportunities for automated information extraction while introducing challenges such as domain-specific terminology, complex sentence structures, and class imbalance. Named Entity Recognition (NER) plays a crucial role in legal text analytics; however, conventional rule-based and generic machine learning approaches often struggle with contextual ambiguity and nested entities. This study proposes a hybrid Legal-BERT-based framework integrated with a semantic similarity filtering mechanism to improve entity boundary precision and extraction reliability. The model leverages domain-adapted contextual embeddings and semantic validation to reduce false positives and enhance prediction consistency. Experimental results demonstrate strong performance, achieving a precision, recall, and F1-score of 0.92 with an overall accuracy of 0.99. A Streamlit-based web application is developed for document upload, entity visualization, and statistical analysis, enabling practical deployment in legal workflows. The proposed framework provides a scalable solution for automated legal entity recognition and intelligent legal text analytics.