Abstract:
This work addresses the challenge of classifying public sector organizations across multiple European languages using only their official names, a critical step for entity disambiguation in knowledge graph population. We employ ontology-based knowledge extraction to evaluate three Natural Language Processing approaches: rule-based keyword extraction, zero-shot Natural Language Inference, and embedding-based semantic similarity —under low-context, low-resource assumptions. Large Language Models are integrated accross all three techniques. Our methodology systematically evaluates multilingual preprocessing, various state-of-the-art models, different supervision regimes, classification structures, and parameter optimization. We conduct a detailed evaluation across three specific domains (healthcare, administration, education) spanning multiple European countries, analyzing performance in relation to lexical structure and class balance.
Results demonstrate that lightweight rule-based methods, particularly TF-IDF keyword selection, are effective in multilingual scenarios with minimal supervision. Natural Language Inference models offer competitive zero-shot performance but show deficiencies with unbalanced class distribution. Embedding-based methods provide the most consistent generalization across languages, with evidence of class coherence in vector space. We apply these techniques to a real-world use case — classifying contracting authorities in the EU Contract Hub platform - and outline additional applications and extensions for governance objectives and ontology refinement. This work highlights the feasibility of ontology-guided multilingual classification from short texts and its contribution to entity disambiguation challenges in formal knowledge representation systems, particularly when integrating diverse European organizational entities into structured knowledge bases.
