Research map / Large language models / Language, safety and interpretability
Diverse Sign Language Translation
Xin Shen, Shen Lei, Shaozu Yuan, Heming Du, Haiyang Sun, Xin Yu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train with shuffled multi-reference strategy instead of Sample One (ablation) | DivSLT stage 1 · PHOENIX14T-Div test | BLEU-BM: 18.43 -> 24.20 | Improved |
| add stage 2 RL training (ablation) | DivSLT · PHOENIX14T-Div test | BLEU-BM: 24.20 -> 25.29 | Improved |
| add stage 2 RL training (ablation) | DivSLT · PHOENIX14T-Div test | pwb: 22.26 -> 26.30 | Hurt |
| use mixed XE+RL loss in stage 1 instead of XE only (ablation) | DivSLT stage 1 · PHOENIX14T-Div test | rfb-BM: 20.74 -> 19.18 | Hurt |
| use RL loss only in stage 2 instead of mixed loss 0.5/0.5 (ablation) | DivSLT stage 2 · PHOENIX14T-Div test | rfb-BM: 21.37 -> 22.12 | Improved |
| use diverse beam search instead of beam search (ablation) | DivSLT · PHOENIX14T-Div test | pwb: 57.14 -> 26.30 | Improved |
| use diverse beam search instead of beam search (ablation) | DivSLT · PHOENIX14T-Div test | rfb-BM: 23.35 -> 22.12 | Hurt |
| apply two-stage multi-reference paradigm | GFSLT · PHOENIX14T-Div test | BLEU-BM: 19.76 -> 25.29 | Improved |
| apply two-stage multi-reference paradigm | RNN · PHOENIX14T-Div test | BLEU-BM: 12.77 -> 13.97 | Improved |
| apply two-stage multi-reference paradigm | Transformer · PHOENIX14T-Div test | BLEU-BM: 10.49 -> 12.31 | Improved |
| apply two-stage multi-reference paradigm | GASLT · PHOENIX14T-Div test | BLEURT-BM: 49.07 -> 55.37 | Improved |
| add DivSLT-tailored vision-language pre-training (ablation) | DivSLT · PHOENIX14T-Div test | BLEU-BM: 25.29 -> 25.79 | Improved |
Closest papers by meaning
- Scaling Sign Language Translation (NeurIPS 2024 · similarity 0.92)
- Sign2GPT: Leveraging Large Language Models for Gloss-Free Sign Language Translation (ICLR 2024 · similarity 0.90)
- LLMs are Good Sign Language Translators (CVPR 2024 · similarity 0.90)
- Lost in Translation, Found in Context: Sign Language Translation with Contextual Cues (CVPR 2025 · similarity 0.89)
- Lost in Translation, Found in Embeddings: Sign Language Translation and Alignment (NeurIPS 2026 · similarity 0.89)
- YouTube-SL-25: A Large-Scale, Open-Domain Multilingual Sign Language Parallel Corpus (ICLR 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
