BackTranslation2.0: A Linguistically Motivated Metric to Assess Sign Language Production

We introduce BackTranslation2.0, a linguistically grounded evaluation metric for text-to-sign translation that moves beyond naïve backtranslation. The approach adopts an agentic framework in which a deterministic pipeline orchestrates a suite of specialised tools to assess four scoring dimensions (grammatical correctness, phonological accuracy, motion fluency, and generation fidelity), with LLM-based cross-referential comparison modules checking consistency across tools against linguistic expectations. Validated on a British Sign Language dataset annotated by native Deaf raters, the metric demonstrates stronger correlation with human judgements than six baseline metrics.

arXiv Preprint  ·  Project Page