Signer dependence can greatly inflate sign language translation scores
Rethinking Sign Language Translation: The Impact of Signer Dependence on Model Evaluation
Computation and LanguageArtificial Intelligence
Summary
Sign language translation models aim to turn sign language videos into written or spoken language, but testing these models often mixes up signers seen in training with those in testing. This causes models to seem much better than they really are, because they learn to recognize specific people's signing styles instead of the language itself. The authors found that when testing on new, unseen signers, model performance drops dramatically. They suggest using testing setups that keep signers separate to get a true measure of how well these models understand sign language in general.
sign language translationsigner dependencedeep learningcross-validationBLEU scoreROUGE scoredataset splitsgeneralizationnatural language processingcomputer vision
Authors
Keren Artiaga, Sabyasachi Kamila, Haithem Afli, Conor Lynch, Mohammed Hasanuzzaman
Abstract
Sign Language Translation has advanced with deep learning, yet evaluations remain largely signer-dependent, with overlapping signers across train/dev/test. This raises concerns about whether models truly generalise or instead rely on signer-specific regularities. We conduct signer-fold cross-validation on GFSLT-VLP, GASLT, and SignCL, three leading, publicly available, gloss-free SLT models, on CSL-Daily and PHOENIX14T. Under signer-independent evaluation, performance drops sharply: on PHOENIX14T, GFSLT-VLP falls from BLEU-4 21.44 to 3.59 and ROUGE-L 42.49 to 11.89; GASLT from 15.74 to 8.26; and SignCL from 22.74 to 3.66. We also observe that in CSL-Daily many target sentences are performed by multiple signers, so common splits can place identical sentences in both training and test, inflating absolute scores by rewarding recall of recurring sentences rather than genuine generalisation. These findings indicate that signer-dependent evaluation can substantially overestimate SLT capability. We recommend: (1) adopting signer-independent protocols to ensure generalisation to unseen signers; (2) restructuring datasets to include explicit signer-independent, sentence-disjoint splits for consistent benchmarking; and (3) reporting both signer-dependent and signer-independent results together with train-test sentence overlap to improve transparency and comparability.