We provide a novel approach for automatically creating high-precision sense-annotated parallel corpora, and put forward a specifically tailored fine-tuning strategy for exploiting these sense annotations during training without introducing any additional requirement at inference time, for improved word sense disambiguation in machine translation.