Towards an In-Context LLM-Based Approach for Automating the Definition of Model Views

Abstract

In the Model-Driven Engineering (MDE) of complex systems, multiple models represent various systems’ aspects. In practice, these models are often unconnected and specified using different modeling languages. Model view solutions can be employed to automatically combine such models. However, writing model view definitions is not trivial. When modeling languages are semantically distant and/or have a large number of concepts, it can quickly become difficult to manually identify the language elements to be selected, associated, or queried to build a model view. As a solution, this paper proposes an in-context Large Language Model (LLM)-based approach to assist engineers in writing model-view definitions. Notably, we rely on LLMs and Prompt Engineering techniques to automatically generate drafts of model-view definitions by providing as input only minimal information on the modeling languages to be combined. We implemented our approach by integrating the EMF Views solution for model views with the LangChain framework for LLM-based applications. To this end, we tailored LangChain to handle EMF metamodels. We validated our approach and implementation on a set of model views originally specified either in VPDL, the ViewPoint Definition Language of EMF Views, or as ATL model-to-model transformations. We compared these original model view definitions with the ones we automatically generated. The obtained results show the feasibility and applicability of our approach.