Rationalization for Explainable NLP: A Survey

Natural language explanations are more accessible than traditional explainability methods: The paper identifies that numerical techniques like LIME and Shapley values, as well as visualization methods like saliency heatmaps, "require specialized knowledge," whereas rationalization—which provides explanations in plain language—is "intuitive, human-comprehensible, and accessible to non-technical users."

There's a significant trade-off between AI performance and explainability: The survey notes that "recent advances in deep learning have improved the performance of many Natural Language Processing tasks" but this improvement "comes at the expense of model explainability," highlighting a key tension in developing responsible AI systems.

Rationalization is an emerging but disorganized field: As the first comprehensive survey of rationalization literature (2007-2022), the paper reveals that while rationalization has become "an attractive technique" due to improvements in natural language generation, the field lacks organization and standardization across different NLP applications.

A new framework called Rational AI (RAI) is needed to advance the field: The authors introduce "Rational AI" as a new subfield within Explainable AI to address gaps in current rationalization approaches and establish more systematic methods for generating trustworthy explanations.

Standardized evaluation methods and resources are lacking: The survey catalogs available methods, evaluations, code, and datasets across various NLP tasks, implying that the field currently lacks consistent standards for assessing whether AI explanations are actually reliable and trustworthy.