Knowledge Graphs for Empirical Concept Retrieval

Lenka Tetková, Teresa Karen Scheidt, Maria Mandrup Fogh, Ellen Marie Gaunby Jørgensen, Finn Årup Nielsen, Lars Kai Hansen

Abstract

Explainable AI (XAI) often relies on human-interpretable concepts to describe model decisions. However, defining these concepts and gathering representative data remains a significant bottleneck. This paper presents a workflow that leverages Knowledge Graphs (KGs) like Wikidata and WordNet to automate concept definition and data retrieval. By grounding concepts in structured human knowledge, we enable personalized explanations and demonstrate that modern AI models naturally align with human hierarchical structures.

Introduction

Concept-based explainability, such as Concept Activation Vectors (TCAV), offers a path toward transparent AI by decomposing model logic into human-friendly terms. Despite its promise, the field faces two primary challenges: data scarcity for niche concepts and the "alignment problem"—ensuring that a model's internal representation of a concept matches human intent. Current methods often rely on small, manually labeled datasets, which are expensive to scale and prone to subjectivity.

Methodology

We propose an interactive workflow that utilizes the semantic richness of Knowledge Graphs to bridge the gap between abstract human concepts and raw empirical data.

Concept Definition and Retrieval

Users interact with a KG to select specific nodes, allowing for precise disambiguation (e.g., distinguishing "Apple" the fruit from the corporation). This hierarchical structure enables users to navigate between broad categories and specific sub-concepts. Once a concept is defined via its unique KG identifier, the system automatically retrieves corresponding images from Wikimedia Commons and text from Wikipedia.

Overview of the concept definition, robustness testing, and alignment evaluation workflow.

Selecting the Right Knowledge Graph

The choice of KG depends on the specific task and the nature of the concepts required. For general image classification, WordNet or Wikidata are often sufficient. For specialized domains like medicine, expert-created KGs are necessary to ensure accuracy.

Decision tree for selecting appropriate Knowledge Graphs based on task requirements and concept types.

Results

The effectiveness of KG-retrieved data was validated across multiple state-of-the-art models, including ViT and data2vec for vision, and BERT and RoBERTa for text.

Performance and Robustness

Concepts defined through our automated workflow performed comparably to, and in some cases better than, those derived from manually labeled datasets like Pascal VOC. Robustness tests showed that explanations remained stable across different random negative sets, suggesting that KG-defined concepts occupy well-delineated regions in the model's latent space.

Hierarchical Alignment

A key finding is the inherent alignment between AI models and human knowledge structures. In "triplet tests," models consistently grouped sub-concepts (e.g., "apple" and "pear") closer to each other than to unrelated concepts, mirroring the hierarchical relationships found in KGs. This suggests that self-supervised learning naturally captures human-like taxonomies in its deeper layers.

Discussion and Conclusion

While successful in later model layers, concept extraction is less reliable in earlier layers where low-level features dominate. Furthermore, broad concepts like "sport" exhibit lower alignment than concrete ones like "motor vehicle" due to their inherent diversity.

By integrating Knowledge Graphs into the XAI pipeline, we provide a formal, scalable route for personalized and robust explanations. Future work will explore the synergy between Large Language Models and KGs to further refine concept definitions and address "common sense" reasoning gaps.

Summary updated

BibTeX
@inproceedings{tetkova2024knowledge,
  author = {Lenka Tetková and Teresa Karen Scheidt and Maria Mandrup Fogh and Ellen Marie Gaunby Jørgensen and Finn Årup Nielsen and Lars Kai Hansen},
  title = {{Knowledge Graphs for Empirical Concept Retrieval}},
  booktitle = {xAI},
  year = {2024},
  doi = {10.1007/978-3-031-63787-2\_9},
}
Figure