Explainable Artificial Intelligence (XAI) has become a critical feature for the responsible use of machine learning (Ml) and deep learning (Dl) systems in sectors with high stakes like the healthcare, engineering, cyber-security and natural language processing (NLP) industries. A review of twenty-nine foundational peer-reviewed and preprint papers that cover different types of XAI, model-agnostic explanation techniques (post-hoc) such as SHAP/LIME, and saliency/attention-based interpretability methods between 2022–2026, focusing on those in the context of text-based and clinical applications. The review concentrates on four thematic areas of focus: (i) conceptual foundations and cross-domain framing approaches for XAI; (ii) Shapley Additive explanations (SHAP) for engineering, environmental, Cyber security, and clinical prediction applications; (iii) Local Interpretable Model-Agnostic Explanations (LIME) for biomedical signals, natural language processing, and image classification; and (iv) saliency maps and attention-based XAI, with a focus on natural language processing and human evaluation for explanation quality. In these clusters, the literature reviewed suggests that post hoc explanation methods have consistent advantages of greater trustability and support by stakeholders for feature level auditing, but significant issues of faithfulness, stability, standardization that do not go away persistently, and inequities of evaluation across tasks and domains. To this end four major research gaps are singled out in the review: a lack of a single and cross-method evaluation protocol, limited representation of complex NLP tasks, lack of domain adaptation for clinical and specialized text and lack of robust human-cantered assessment of the utility of explanations.The gaps drive further research into (easily) adaptable, accurate, and comprehensible explanation frameworks, especially in high stakes and regulated natural language-based AI systems. 

Explainable AI (XAI), Model-Agnostic Techniques, Saliency Maps, Interpretability, Machine Learning Review, Cross-Domain Application