Explainable AI in Legal Text Analysis: A Comprehensive Review of Legal Threat Assessment Approaches

Abstract

The volume of legal documents continues to expand rapidly, and their language is becoming increasingly intricate, making manual review both time-intensive and vulnerable to individual bias and inconsistency. As a result, recent research has increasingly focused on the use of machine learning techniques to predict threat levels and evaluate legal risks directly from textual data. Although such models have demonstrated promising predictive performance, a major limitation remains the lack of transparency in their decision-making processes. In legal practice, where accountability and justification are essential, predictions without clear explanations are difficult to accept and apply. This review examines studies published between 2018 and 2025 addressing automated legal document analysis, with particular emphasis on methods that combine accurate threat prediction with explainability. The reviewed works cover a broad range of approaches, including natural language processing, deep learning, transformer-based architectures, graph-based models, large language models, and contract analysis techniques. Explainable artificial intelligence methods such as LIME, SHAP, and attention-based visualization are examined for their role in making model predictions transparent and legally meaningful. Recurring design patterns observed across the reviewed systems are synthesized into a conceptual framework of six sequential stages covering input, text processing, feature encoding, threat classification, explainability, and output. The analysis identifies dataset scarcity, inconsistent evaluation practices, and limited generalization across legal domains as the primary barriers to practical deployment. The review concludes that explainability is a foundational requirement rather than an optional feature for any AI system intended to support legal decision-making.