Lack of structural correspondence in unsupervised datasets leads to distorted translations. Preserving semantic relations ensures high-fidelity output across disparate domains.