Vision-language Models (VLMs) can generate natural language explanations for complex decision-making scenarios. However, these explanations can sometimes be misleading, sounding convincing while containing incorrect or unsupported information. Automated explanation-quality metrics help researchers and practitioners evaluate large collections of explanations, but their scores alone provide limited support for identifying failure cases and understanding the reliability of these metrics. This paper presents a case study on how interactive visualizations can help researchers and practitioners evaluate both AI-generated explanations and automated explanation-quality metrics across large datasets. Based on formative interviews with nine AI researchers and practitioners, we designed an interactive visualization tool named ExCritique, situated in the task of Visual Question Answering (VQA). ExCritique visualizes the distributions of multiple automated quality metrics, allows users to explore thousands of data items, filter items and inspect the evidence behind individual scores. We compared ExCritique with a baseline interface without interactive visualizations in a controlled study involving 16 AI researchers and practitioners. Our findings suggest that the interactive visualizations in ExCritique helped users identify explanation-quality issues more efficiently and better understand automated quality metrics, supporting an effective way to measure the reliability of AI-generated explanations.

翻译：暂无翻译