Target-Side Paraphrase Augmentation for Sign Language Translation with Large Language Models

The article presents a novel target-side paraphrase augmentation strategy for sign language translation (SLT) using GPT-4o to generate paraphrased variants of spoken-language sentences while keeping sign inputs constant. A Signformer-style Transformer is trained in a two-stage process involving pre-training on the augmented dataset and fine-tuning on original references, leading to a BLEU-4 score improvement on the PHOENIX14T dataset from 9.56 to 10.33. This research highlights the potential of LLM-generated paraphrases to enhance decoder generalization in SLT, emphasizing the need for semantic evaluations alongside traditional lexical metrics.