STT-CL: A Unified Spatio-Temporal-Task Framework for Continual Learning in Gloss-Free Multilingual Sign Language Translation

Sign Language Translation (SLT) is fundamentally a process of intersemiotic translation, involving the transfer of meaning between visual-gestural and spoken language systems. However, most existing SLT studies focus on static bilingual settings, overlooking the dynamic and multilingual nature of real-world deployment. Moving towards Continual Learning in Multilingual Sign Language Translation (CL-MSLT) is essential but remains challenging due to catastrophic forgetting, spatio-temporal redundancy, and cross-lingual heterogeneity. To overcome these issues, we propose STT-CL, a Spatio-Temporal-Task unified framework operating in a two-stage paradigm. First, we establish a noise-robust visual foundation by embedding the proposed Multi-Granularity Spatio-Temporal Dynamic Calibration module into the visual backbone, effectively filtering out non-informative frames during pre-training. Second, to adapt the frozen backbone to heterogeneous sign languages without forgetting, we introduce Language-Guided Visual Prompt Tuning for task-aware context injection and Spatio-Temporal Decoupled Mixture-of-Experts for physical parameter isolation. Drawing upon the Inhibitory Control theory in cognitive translatology, this design structurally segregates parameters to suppress retroactive interference, ensuring the retention of historical knowledge. To facilitate research in this domain, we curate the CL-MSLT benchmark based on the SP-10 dataset with rigorous sequential protocols. Extensive experiments demonstrate that STT-CL achieves state-of-the-art performance, effectively mitigating forgetting while maintaining high translation quality across diverse task permutations.