TRIZ-Aware Regularization for Invariant-Preserving Neural Machine Translation in Metrological and Scientific Domains (v1.0)
Year: 2025
Type: article
Abstract: This release presents TRIZ-Aware Loss v1.0, a semantic regularization framework for invariant-preserving neural machine translation in metrological and scientific domains. The approach integrates penalties for placeholder, formula, and constant mismatches into the training objective, ensuring structural and semantic stability during translation. Implemented on the Helsinki-NLP/opus-mt-en-uk model with LoRA fine-tuning, it demonstrates improved retention of mathematical invariants and placeholder consistency across the A-128 corpus and the External-50 evaluation subset. The dataset and scripts define a reproducible baseline for TRIZ-driven explainable translation systems, aligning linguistic fluency with formal domain integrity. This version marks the first public release of the TRIZ-Aware Loss framework (v1.0) for Corpus A on Zenodo. © 2025 Yurii Kuzmenko, Olena Trofimova.For academic use only. Redistribution and commercial use prohibited.
Authors: Kuzmenko, Yurii, Trofimova, Olena
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7095040367
FWCI:
Cites:
Cited by:
Related to: