Rationalization for explainable NLP: a survey.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Explainable AI

Background:

• Deep learning advances in NLP tasks improve performance but reduce model explainability.
• Black-box models hinder understanding of system internals and decision-making processes.
• Existing explainability methods (e.g., LIME, Shapley, saliency heatmaps) require specialized knowledge and are insufficient.

Purpose of the Study:

• To provide the first comprehensive survey of rationalization literature in NLP from 2007 to 2022.
• To organize and analyze existing methods, evaluations, code, and datasets in NLP rationalization.
• To introduce Rational AI (RAI) as a new subfield of Explainable AI to advance rationalization.

Main Methods:

• Systematic literature review of rationalization techniques in Natural Language Processing.
• Analysis of methods, evaluation metrics, available code, and datasets.
• Categorization of research and identification of trends and gaps.

Main Results:

• Rationalization provides intuitive, human-comprehensible natural language explanations (rationales) for model outputs.
• The field of rationalization in NLP is currently disorganized.
• Identified key methods, evaluation strategies, and resources used in rationalization research.

Conclusions:

• Rationalization is an accessible explainability technique for non-technical users.
• The survey establishes a foundation for the emerging field of Rational AI (RAI).
• Future research directions and opportunities in NLP rationalization are discussed.