Abstract
Accurate communication between deaf and hearing individuals remains challenging because qualified sign-language interpreters are not always available, especially for under-resourced sign languages such as Syrian Arabic Sign Language. This paper presents a newly developedpreliminary dataset for Syrian Arabic Sign Language and establishes a baseline end-to-end system for translating continuous sign language videos into grammatically correct written Arabic. The proposed system provides an initial benchmark for video-to-text sign language translation and serves as a foundation for future research on Syrian Arabic Sign Language translation. The task is formulated as video-to-text sign language translation. To capture fine-grained signing cues, including hand configurations, motion dynamics, and facial expressions, we combine features extracted from two pretrained vision models and model their temporaldependencies with a Transformer encoder. An AraBART decoder then generates Arabic sentences conditioned on the encoded video representation. To improve signer invariance, the proposed system incorporates adversarial training through a Gradient Reversal Layer (GRL), while backbone-specific preprocessing is applied to better accommodate the input characteristics of the two visual encoders. The system is evaluated on the newly collected dataset under a signer-dependent setting. The best configuration, incorporating GRL and backbone-specific preprocessing, achieves a WER of 0.166. These results indicate that combining pretrained visual representations, adversarial training, backbone-specificpreprocessing, and AraBART-based text generation is a promising direction for Syrian Arabic Sign Language translation under controlled conditions. However, performance drops substantially in the signer independent setting, indicating that cross-signer generalization remains challenging.