Managing human-AI collaborations within Industry 5.0 scenarios via knowledge graphs: key challenges and lessons learned

• DOI: 10.3389/frai.2024.1247712
• Publication Date: 2024-11-11T06:11:40Z
• Authors: Franz Krause; Heiko Paulheim; Elmar Kiesling; Kabul Kurniawan; Maria Chiara Leva; Hector Diego Estr...; Gernot Stübl; Nazim Kemal Üre; Javier Dominguez-...; Maqbool Khan; Pedro Demolder; Hans Gaux; Bernhard Heinzl; Thomas Hoch; Jorge Martinez-Gil; Agastya Silvina; Bernhard A. Moser
• Keywords: knowledge integration; human-AI collaboration; QA75.5-76.95; 02 engineering and technology; 004; Industry 5.0; semantic web; knowledge graph; Artificial Intelligence; Electronic computers. Computer science; 0202 electrical engineering, electronic engineering, information engineering; process modeling

In this paper, we discuss technologies and approaches based on Knowledge Graphs (KGs) that enable the management of inline human interventions in AI-assisted manufacturing processes in Industry 5.0 under potentially changing conditions in order to maintain or improve the overall system performance. Whereas KG-based systems are commonly based on a static view with their structure fixed at design time, we argue that the dynamic challenge of inline Human-AI (H-AI) collaboration in industrial settings calls for a late shaping design principle. In contrast to early shaping, which determines the system's behavior at design time in a fine granular manner, late shaping is a coarse-to-fine approach that leaves more space for fine-tuning, adaptation and integration of human intelligence at runtime. In this context we discuss approaches and lessons learned from the European manufacturing project Teaming.AI, https://www.teamingai-project.eu/, addressing general challenges like the modeling of domain expertise with particular focus on vertical knowledge integration, as well as challenges linked to an industrial KG of choice, such as its dynamic population and the late shaping of KG embeddings as the foundation of relational machine learning models which have emerged as an effective tool for exploiting graph-structured data to infer new insights.