An Agentic LLM-based Environment for Interactive Domain Knowledge Elicitation and Metamodeling

Abstract

Domain modeling is a knowledge-intensive activity that requires domain experts to iteratively refine abstract representations of their problem space. In practice, eliciting a complete and correct metamodel from textual descriptions expressed in natural language is error-prone and requires multiple human interactions before obtaining satisfactory results. With their recent progress in the state-of-the-art, LLM-based agents offer promise for assisting humans in this process. Existing LLM-based approaches often rely on single-shot generation without any associated validation support. They sometimes provide iterative workflows but lack interaction capabilities with humans, as well as automated conformance checking. As a result, structural errors may remain undetected, and domain knowledge is only partially utilized. They also generally assume modeling expertise on the human side, thus excluding users who cannot properly interpret or correct the generated metamodels. In this paper, we present an agentic LLM-based environment for interactive domain knowledge elicitation and metamodeling. It guides users through iterative steps, validating each step via model conformance checking combined with user validation and document coverage. We propose building metamodels fragment by fragment to progressively uncover structural and conceptual issues. We also support users with different levels of domain and modeling expertise by automatically producing fix suggestions. A user study is planned as future work.