Model size is not the decisive standard: study calibrates inference efficiency in ontology building and knowledge extraction

A rigorous research study by computer science and artificial intelligence researchers showed that increasing the size of large language models does not necessarily entail an automatic or uniform improvement in ontology-building and specialized knowledge-structure extraction tasks. The published paper included a controlled evaluation of thirteen models, ranging from dense-model architectures to mixture-of-experts models, and incorporated open-source Qwen lineages together with closed-source GPT models, aiming to track the actual impact of parameter count on data management and knowledge graph construction.

The researchers based their comparison on a unified architecture for retrieval-augmented generation techniques, using the processing pipeline known as OntoLearner, with a fixed embedding model and search-retrieval settings, standardized command templates, decoding configurations, evaluation metrics and datasets across four domain-specific ontologies in biomedicine and materials science and engineering. The test targeted three core tasks: term-type identification, hierarchical classification discovery, and extraction of non-hierarchical relations between concepts.

The results showed that increasing the number of parameters primarily improves precision rather than recall to the same extent, with the largest gains in dense models ranging from nine billion to twenty-seven billion parameters.Experiments indicated that a dense model with twenty-seven billion parameters outperformed much larger mixture-of-experts models on the term-type identification task, whereas the larger mixture-of-experts models secured superiority in hierarchical classification discovery among open-weight models.

In the domain of extracting non-hierarchical relations between knowledge entities, the task remained extremely difficult for all models regardless of size or lineage, particularly in the materials data science ontology. Fine-grained comparisons between architecture-identical closed-source commercial versions revealed that the model lineage and its internal engineering design often tip the balance more than a mere increase in nominal parameters, making size alone a limited criterion when selecting models to support ontology engineering.

This scientific finding imposes a new practical reality on data teams and software infrastructure in the Gulf, Egypt and the Levant, especially in energy, engineering and health-service sectors that are building local knowledge graphs to document their assets and sensitive data. There is no longer a technical justification for bearing massive inference costs or relying exclusively on giant cloud-model APIs; instead, these results enable companies to run specialized open-source models of twenty-seven billion parameters locally within their national data centers, achieving high extraction accuracy and disciplined compute costs that ensure data sovereignty and operational independence.