A Semantic Approach for Linked Model, Data, and Dataspace Cards

In artificial intelligence, the significance of thorough documentation of models and datasets for publication is underestimated. However, due to the rising trend in the explainability and fairness of AI models, frameworks like Model Cards, Service Cards and Data Cards have emerged to facilitate understanding and reusing those models and datasets. Moreover, the Dataspace concept integrates these resources into Dataspace Cards, a comprehensive framework that systematically captures and organises crucial information to guide model and data selection for a specific application. This paper advocates a Semantic Web approach for transforming Model/Data Cards into Linked Data or knowledge graphs within a Dataspace, rendering them machine-readable and interoperable. A significant contribution is the development of a vocabulary that unifies Data, Model and Dataspace Card ontologies, enhancing consistent documentation and understanding of the Dataspace design. The paper further demonstrates the applicability of the proposed schema in various use cases, including bias detection in BERT-base-uncased and Large Language Models. Additionally, we propose a conceptual semantic approach, examined in-depth for sentiment and emotion analysis, to highlight how extended Dataspace Cards can improve applicability and outcomes. We found that this unified, ontology-driven approach results in more consistent metadata linking and more fine-grained bias detection in BERT-based-uncased than standalone documentation tools relying solely on Model or Data Cards. Furthermore, compared to existing frameworks, the richer interlinking capabilities of our proposed Dataspace Cards also facilitated easier traceability of performance outcomes, thereby ultimately fostering higher trustworthiness and reusability of AI resources.