Sign language translation (SLT) is a challenging task due to the scarcity of labeled data and the heavy-tailed distribution of sign language vocabularies. In this paper, we explore a novel data augmentation approach for SLT: using a large language model (LLM) to generate paraphrases of the target language sentences in the training data. We experiment with a Transformer-based SLT model (Signformer) on three datasets spanning German, Greek, and Argentinian Sign Languages. For models trained with augmentation, we adopt a two-stage regime: pre-train on the LLM-augmented corpus and then fine-tune on the original, non-augmented training set. Our augmented training sets, expanded with GPT-4-generated paraphrases, yield mixed results. On a medium-scale German SL corpus (PHOENIX14T), LLM augmentation improves BLEU-4 from 9.56 to 10.33. In contrast, a small-vocabulary Greek SL dataset with a near-perfect baseline (94.38 BLEU) sees a slight drop to 92.22 BLEU, and a complex Argentinian SL corpus with a long-tail vocabulary distribution remains around 1.2 BLEU despite augmentation. We analyze these outcomes in relation to each dataset’s complexity and token frequency distribution, finding that LLM-based augmentation is more beneficial when the dataset contains a richer vocabulary and many infrequent tokens. To our knowledge, this work is the first to apply LLM paraphrasing to SLT, and we discuss these results with respect to prior data augmentation efforts in sign language translation.

Sign languages are highly diverse across countries and regions, yet most Sign Language Translation (SLT) work remains monolingual. We explore a unified, multi-target SLT model trained jointly on four sign languages (German, Greek, Argentinian, Indian) using a standardized data layer. Our model operates on pose keypoints extracted with MediaPipe, yielding a lightweight and dataset-agnostic representation that is less sensitive to backgrounds, clothing, cameras, or signer identity while retaining motion and configuration cues. On RWTH-PHOENIX-Weather 2014T, Greek Sign Language Dataset, LSA-T, and ISLTranslate, naive joint training under a fully shared parameterization performs worse than monolingual baselines; however, a simple two-stage schedule: multilingual pre-training followed by a short language-specific fine-tuning, recovers and surpasses monolingual results on three datasets (PHOENIX14T: +0.15 BLEU-4; GSL: +0.74; ISL: +0.10) while narrowing the gap on the most challenging corpus (LSA-T: -0.24 vs. monolingual). Scores span from BLEU-4≈ 1 on open-domain news (LSA-T) to >90 on constrained curricula (GSL), highlighting the role of dataset complexity. We release our code to facilitate training and evaluation of multilingual SLT models.