ACL 2026 Drops Bombshell: 120 Sign-Language Datasets, 35 Languages — and a Massive Annotation Problem

Sign-language AI has long been held back by scattered datasets, inconsistent annotations, and limited linguistic coverage—making it hard to build models that work in the real world. A new survey from Yiming Ni, Zhi-Qi Cheng, Jiayu Li, and Wei Cheng, accepted to ACL 2026, takes a massive step toward fixing that. The team cataloged 120 sign-language datasets spanning 35 different sign languages, from widely used ones like American Sign Language to underrepresented variants. They analyze critical pain points: modality imbalance (most datasets focus only on video, ignoring motion capture or gloss annotations), annotation granularity (some offer frame-level labels, others only phrase-level), and signer bias (models trained on a few signers fail on new signers).

To bring order to the chaos, the authors propose a 24-field Sign-Language Datasheet—a structured, reproducible documentation standard for any new dataset. They also released a public GitHub repository that tracks existing datasets, their properties, and evaluation protocols. The survey itself is a practical foundation: researchers can now quickly identify which datasets fit their task (recognition, translation, or production), understand annotation trade-offs, and apply the datasheet to future collection efforts. For the AI community, this means faster iteration, better reproducibility, and more robust sign-language models that actually serve the Deaf and Hard-of-Hearing communities at scale.