Teaching computers to translate between sign languages directly

Direct sign-to-sign translation has been impossible due to lack of parallel corpora—existing systems chain together sign-to-text, text-to-text, and text-to-sign steps, losing visual information and introducing errors. Researchers used back-translation to create synthetic ASL-CSL-DGS pairs from unpaired sign videos and text, then trained a single model for both text-to-sign and direct sign-to-sign translation. The direct approach outperformed cascaded baselines by 20% on motion metrics and 50% on BLEU while running 2.3× faster, offering deaf and hard-of-hearing communities a path to cross-language communication without interpreters.