Semantic Interoperability of Heterogeneous Knowledge Models in Agricultural Advisory Systems: From Expressiveness Assessment to Integration (preprint)

Preprint. This manuscript has not been peer reviewed. It is formatted according to the guidelines of the journal "Actual Problems of Automation and Information Technologies"; the main text is in English, with a title, abstract and keywords in Ukrainian. Purpose. To turn an assessment of the expressiveness of knowledge representation technologies into concrete integration decisions for an agricultural advisory system whose modules – an OWL 2 ontology, a semantic wiki, a knowledge graph built from official registers and a large language model – exchange knowledge in RDF, and to measure what each decision changes in the knowledge that finally reaches the advisor. Methods. Each link of the chain from source data to the advisor's answer is described by the expressiveness profile of the receiving technology, which gives every construct type one of four statuses (preserved, encoded by convention, recovered by query, lost); for a chain of links the weakest status is taken. The profile is complemented by a validity condition for product registrations, role projections of user profiles and an acceptance condition for language model answers. The approach was implemented in Python and Semantic MediaWiki and evaluated in seven experiments on the ALE environment: 253 ontology axioms, 11 336 registered uses of plant protection products linked to the State Register of Ukraine for 2008–2024, up to 22 672 uses in the wiki, and 130 queries answered by two local models under eight context conditions. Results. Closed-world checks in the wiki and open-world inference of the reasoner caught complementary errors in faulty data, so both are needed. A subquery over one n-ary node kept precision at 1.0 at every data volume, whereas a conjunction of property chains stayed at 0.62–0.71, with median query times of 102–125 ms. Between the handbook date and the end of 2024 the number of valid registered uses fell from 9 672 to 8 503, and 18–22% of the products recommended with the n-ary context were no longer registered at the end of 2024; a register-aware check detected all such recommendations. Role projections exposed no private property to other users, advisors or cloud models. With the n-ary context the share of erroneous answers fell from 49–99% under text or fact retrieval to 7–14%, and to 6% after validation, while the median answer time was halved compared with fact retrieval. Conclusions. The expressiveness profile can serve as an integration specification: for each construct and each link it states whether to rely on the receiver, check data at entry, use a query template, route the query to the reasoner or block the answer, and it shows where data validity and privacy must be enforced. Data and code: experiments E1–E7 – 10.5281/zenodo.23448604; ALE ontology – 10.5281/zenodo.23643625; competency questions D8 – 10.5281/zenodo.23878449; context fragments D6 – 10.5281/zenodo.23880420; register sample D12 – 10.5281/zenodo.23878466. Funding: Ministry of Education and Science of Ukraine, project "Development of an Intelligent Support Environment for Agricultural Advisory Services Based on the Integration of Semantic Wiki Technologies and Artificial Intelligence" (state registration No. 0126U000765). Project website: https://www.tsatu.edu.ua/ale/.