Annotation

Annotation: The paper investigates the influence of clarifying instructions and semantic feedback on the quality of alignment between program models of standardization objects generated by large language models. The problem is that independently generated models may be syntactically correct and semantically close, yet difficult to align because of differences in the naming of entities, roles, states, relationships, and units of measurement. Unlike the preliminary validation on a small number of documents, the present study was conducted on a corpus of 32 RFC documents, six generative models, and five independent iterations for each combination of document, model, and scenario. In total, 2,880 models were processed; 14,400 pairs were formed from them, and 85,400 alignment operations were performed using two COMA-like methods: valentine and bdikit. The results show that general instructions on semantic naming and code structuring improve the relative alignment quality by 46.2%, whereas semantic feedback worsens the result, improving quality by only 40.3% relative to the control condition. This finding should be taken into account when developing tools for formal model generation, static analysis, model alignment, and the preparation of intermediate representations for the analysis of technical and regulatory documents.