A conversational tool for AI-driven feedback for UML modeling

Abstract Teaching the unified modeling language (UML) is a critical task that can benefit from emerging artificial intelligence-based solutions. This paper presents UML Miner, a visual paradigm plugin designed to support the teaching and learning of UML by capturing and analyzing students’ modeling activities to deliver personalized, context-aware feedback. The tool records fine-grained software modeling actions, storing them in an event log. For each diagram, a natural language description and its XML representation are stored separately. Modeling behavior is reconstructed through process mining techniques and compared with a reference model created by the instructor using conformance checking. This analysis adopts a declarative approach, leveraging the declare language to specify behavioral constraints without enforcing a rigid action sequence, thereby accommodating multiple valid modeling strategies while preserving semantic rigor. The detected violations, together with the diagram’s natural language and XML representations, are integrated into a retrieval-augmented generation-enabled large language model, which combines accumulated knowledge with the discovered process constraints to generate enriched, pedagogically meaningful feedback in real time. An exploratory case study conducted in a real-world learning scenario evaluates the tool’s performance from both qualitative and quantitative perspectives, demonstrating its potential to improve learning outcomes and student engagement.