NLP-Based Semantic Matching on ECLASS: Design and Validation of an Industrie 4.0 Matching Service

Abstract

Semantic interoperability remains a key challenge in industrial data exchange, particularly with respect to the shared understanding of heterogeneous term definitions, even within a single term resource like ECLASS. This contribution presents a Natural Language Processing (NLP)-based Semantic Matching Service that derives semantic similarity relations from textual definitions of ECLASS Classification Classes using embedding models. Using the term elements IRDI, preferred name and definition, textual descriptions are normalised, embedded into a high-dimensional vector space and compared by means of angular similarity to obtain similarity scores in the interval [0, 1]. These scores are interpreted as weighted edges in a term similarity graph, in which direct and transitive relations are modelled within a fuzzy-graph framework. To ensure consistency between direct and inferred similarities, the model enforces max-product transitivity, requiring that indirect similarities do not exceed their corresponding direct relations. Evaluation on selected ECLASS Segments shows semantically plausible direct and transitive matches as well as identifiable local similarity clusters, while also revealing exponential similarity decay along transitive paths and scalability limits of densely connected graphs. The findings underline the heuristic character of similarity scores and their dependence on the quality of the underlying textual definitions.