Knowledge graphs for empirical concept retrieval

Knowledge graphs can help make AI explanations more personalized: The paper demonstrates that using knowledge graphs (like Wikidata or WordNet) allows users to define concepts interactively rather than relying on pre-set formal definitions, enabling explainability methods to be tailored to individual user needs and intentions.

This approach bridges human understanding and AI understanding: The research found "good alignment between the models' representations of concepts and the structure of knowledge graphs, i.e., human representations," suggesting that AI systems trained with knowledge graph-based concepts better match how humans actually think about and organize information.

Knowledge graphs solve a practical problem in concept-based explainability: The paper addresses a real challenge in the field—that while concept-based methods like TCAV are appealing because they avoid complex formal definitions, it has been difficult to create relevant concept datasets. Knowledge graphs provide a systematic way to solve this problem.

The method works reliably across different domains and explanation techniques: The researchers tested their approach on two different concept-based explainability methods (CAVs and CARs) and found that concept datasets derived from knowledge graphs "provide robust and accurate explanations," demonstrating the approach's versatility.