Unified Metrology–TRIZ Neural Machine Translation (UMT–NMT) v1.0: Proof-of-Concept Study

This record presents UMT–NMT v1.0, a domain-specific neural machine translation (NMT) proof-of-concept for English→Ukrainian translation of normative documents in metrology. It introduces Corpus A (128 aligned sentence pairs, including 50 formula-focused pairs) and a TRIZ-aware loss function that encourages the preservation of structural elements such as placeholders, formulas, and physical constants. The proposed loss integrates functional analysis and contradiction resolution principles from TRIZ to balance structural fidelity with lexical adequacy. Evaluation on the current dataset confirms full preservation of placeholders and formula tokens (PPR = 1.0, FTA = 1.0) and a BLEU score of 30.09, reflecting limited fluency due to the small data size. General-purpose NMT models often distort domain-critical elements, such as replacing J·s with J, which can invalidate measurement results in metrology. This release addresses such risks by prioritizing structural accuracy under low-resource conditions. This version includes: the aligned sentence pairs (Corpus A), placeholder annotations and TRIZ semantic tags, supporting formula and token graphs (Graphviz, PNG), a custom SentencePiece tokenizer, adapted to preserve domain tokens. Revised sub-corpus: The file CORE_CR_formula50_Zenodo.csv supersedes the earlier CORE_CR_50_formula_Zenodo.csv from Dataset v1.0. Future work (v2.0) will explore semantic bridges between classical and quantum metrology and extend the TRIZ-inspired framework to deeper abstraction layers. © 2025 Yurii Kuzmenko, Olena Trofimova. For academic use only. Redistribution and commercial use prohibited.