A Comprehensive Analysis of XNLP Techniques, Applications, and Evaluation Methods
Preprint · arXiv:2502.00837
Natural Language Processing (NLP) is now embedded in critical sectors including healthcare, finance, and customer relationship management, where models such as GPT-4o, Gemini, and BERT increasingly inform decisions. This survey examines explainable NLP (XNLP) as it is actually deployed, working through seven application domains: medicine, finance, systematic reviews, CRM, chatbots, social and behavioral science, and human resources. For each domain, we ask what kind of explanation the setting needs, which methods are used there, and how they are evaluated.
A structured cross-domain synthesis contrasts how those requirements diverge. We compare the main explanation method families on scope, evidence of faithfulness, and computational cost, and we propose a two-tier evaluation protocol that separates a shared technical core of metrics from the domain-specific validation layer through which those metrics have to be read. The survey closes with research directions, among them personalized explanations, human-in-the-loop evaluation, and mechanistic interpretability for large language models.
Key statistics from our comprehensive literature review
Hierarchical classification of explainability techniques in NLP
Interactive visualizations of our survey findings
Our cross-domain analysis reveals critical insights for researchers and practitioners deploying XNLP systems.
Explainability is not one-size-fits-all. Medical applications prioritize clinical validity and patient safety, while financial applications emphasize regulatory compliance and adversarial robustness. Each domain requires tailored explanation strategies.
A significant disconnect exists between technical metrics (fidelity, faithfulness) and practical utility. The M⁴ benchmark reveals that explanation methods perform inconsistently across modalities, highlighting the need for domain-specific validation protocols.
Large language models present unique interpretability challenges. Chain-of-thought explanations may not faithfully reflect internal reasoning. Mechanistic interpretability through sparse autoencoders shows promise but remains incomplete.
Empirical evidence shows transparency need not cost accuracy: RETAIN reaches an AUC of 0.8705 for heart-failure prediction, on par with an unconstrained RNN, while staying auditable; bias-assessment tools agree with human reviewers (Cohen’s kappa 0.42–0.48). Sources are cited in the paper.
Comparison of XNLP requirements, methods, and challenges across seven major application domains.
| Domain | Primary Need | Key Methods | Unique Challenges |
|---|---|---|---|
| Medicine | Clinical actionability; patient safety | RETAIN, SHAP, LIME, attention visualization | HIPAA/GDPR constraints; clinical workflow integration |
| Finance | Regulatory compliance; fraud detection | SHAP, LIME, attention heatmaps | Adversarial gaming; real-time constraints |
| CRM | User satisfaction; personalization | Attention visualization, rationale generation | Multi-language requirements; privacy concerns |
| HR | Fair hiring; bias reduction | Counterfactual explanations, feature importance | Legal liability; cross-cultural validity |
| Social Science | Content moderation; safety | SHAP, LIME, attention-based methods | Cultural sensitivity; annotation bias |
| Systematic Reviews | Reproducibility; efficiency | Rule-based extraction, active learning | Heterogeneous corpora; tool integration |
| Chatbots | User trust; understanding | Dialogue explanations, attention visualization | Real-time generation; conversational flow |
Curated collection of tools, datasets, and supplementary materials from our survey.
Hierarchical classification of techniques, applications, and evaluation methods in JSON format.
View Taxonomy →Curated links to LIME, SHAP, Captum, HateXplain, ERASER, and more.
View Resources →If you find this survey useful in your research, please cite our work.
@misc{mohammadi2025explainability,
title={Explainability in Practice: A Survey of Explainable {NLP}
Across Various Domains},
author={Mohammadi, Hadi and Bagheri, Robert A. and
Giachanou, Anastasia and Oberski, Daniel L.},
year={2025},
eprint={2502.00837},
archivePrefix={arXiv},
url={https://arxiv.org/abs/2502.00837}
}
See CITATION.cff for machine-readable metadata.