Beyond BLEU: Linguistic Invisibility and Interactional Repair Sequence in End-to-End Sign Language Translation

Abstract

Recent advances in end-to-end sign language translation (SLT) have achieved benchmark performance, yet little is known about whether these systems preserve the multi-channel linguistic structures that are essential for real-world communication. We argue that current optimization and evaluation practices create a form of linguistic invisibility, where interactionally decisive non-manual signals (NMS) are systematically underrepresented despite high translation scores.To empirically examine this issue, we analyze an interactional repair sequence from a Japanese Sign Language (JSL) conversational corpus as a diagnostic probe. Combining qualitative interactional analysis with kinematic measurements, we demonstrate a consistent manual–mouth decoupling pattern in which semantic resolution is carried primarily by mouthing while manual articulation remains largely constant. We show that such cross-channel contrast is unlikely to be preserved under current end-to-end training objectives that prioritize global motion similarity. Based on these findings, we argue that progress in SLT should be evaluated not only by sequence-level accuracy but also by the preservation of linguistically contrastive structures, motivating the development of diagnostic, multi-channel evaluation protocols for future SLT benchmarks. We therefore propose incorporating multi-channel diagnostic evaluation sets and decoupling-sensitive metrics into future SLT benchmarking frameworks, providing a pathway toward models that achieve both high performance and linguistic structural visibility.

Resources

Details