On the Importance and Evaluation of Narrativity in Natural Language AI Explanations

Current AI explanations often fail to explain "why": Most existing explanation methods provide static lists of important features, which tell users what influences a prediction but don't explain the underlying why—a critical gap the paper identifies through analysis of Natural Language Generation approaches to explainability.

Narrative structure matters for human understanding: Drawing on social science and linguistics research, the paper argues that effective explanations should have four key properties: continuous structure, cause-effect mechanisms, linguistic fluency, and lexical diversity. These narrative qualities help people actually comprehend AI decisions rather than just receive feature lists.

Standard metrics for evaluating explanations are inadequate: The research demonstrates that common Natural Language Processing evaluation metrics (based on token probability or word frequency) fail to capture whether explanations are truly informative, and can even rate meaningless "tautological text" as highly as genuine explanations.

New metrics needed to assess explanation quality: The authors propose seven automatic metrics specifically designed to measure narrative quality across the four identified dimensions, showing these new metrics better distinguish between truly descriptive explanations and those that merely sound fluent without conveying real understanding.

Existing AI explanation methods underperform on narrative quality: When benchmarked against six datasets, current state-of-the-art explanation generation methods score poorly on the proposed narrative metrics, suggesting significant room for improvement in how AI systems communicate their reasoning to users.