On the Evaluation of the Plausibility and Faithfulness of Sentiment Analysis Explanations

Explainability methods need standardized evaluation: The paper identifies a critical gap in how AI explanations are currently assessed across different studies and datasets, highlighting that the lack of common evaluation frameworks is slowing progress in making AI systems more transparent and trustworthy.

Not all explanation methods are equally reliable: The research found that "anchor" explanations performed better at aligning with human judgment compared to other methods, suggesting that organizations should be selective about which explainability techniques they adopt for sentiment analysis systems.

Explanation quality depends on task complexity: The study demonstrates that more complex reasoning tasks make it significantly harder for AI explanation methods to identify the actual evidence supporting their decisions, meaning users should have lower confidence in explanations for nuanced or ambiguous sentiment cases.

AI architecture matters for transparency: The paper shows that different underlying model architectures (particularly transformers versus convolutional networks) produce substantially different explanation results, indicating that choices about how to build AI systems have real implications for whether those systems can be meaningfully explained to humans.

Human agreement is a crucial test: By measuring whether AI explanations match human judgment on a custom dataset, the research establishes that real human validation—not just technical metrics—should be part of evaluating whether explanations are actually trustworthy and useful.