Research map / Large language models / Language, safety and interpretability
SIGNET: Motion-Level Knowledge Transfer for Cross-Language Sign Language Translation
Sobhan Asasi, Ozge Mercanoglu Sincan, Richard Bowden
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 |
|---|---|---|---|
| Freeze pretrained backbone and adapt with light modules instead of fine-tuning (ablation) | CSL-News pretrained backbone · Phoenix14T (cross-lingual) | BLEU-4: 11.47 -> 23.30 | Improved |
| Freeze pretrained backbone, same-language transfer (ablation) | CSL-News pretrained backbone · CSL-Daily | BLEU-4: 25.56 -> 24.65 | No clear effect |
| Freeze pretrained backbone (YT-ASL) (ablation) | YT-ASL pretrained backbone · CSL-Daily | BLEU-4: 13.8 -> 20.7 | Improved |
| Restrict aggregation input to hand features rather than adding body and face (ablation) | SIGNET · Phoenix14T | BLEU-4: 25.12 -> 27.78 | Improved |
| Use top-2 gating instead of using all 3 experts (ablation) | SIGNET (3 backbones) · Phoenix14T | BLEU-4: 26.30 -> 27.78 | Improved |
| Use top-2 instead of top-1 gating (ablation) | SIGNET (3 backbones) · Phoenix14T | BLEU-4: 27.20 -> 27.78 | Improved |
| Add Stage II contrastive alignment (ablation) | SIGNET · Phoenix14T | BLEU-4: 27.63 -> 27.78 | Improved |
| Add Stage II contrastive alignment (ablation) | SIGNET · CSL-Daily | BLEU-4: 28.27 -> 26.50 | Hurt |
| Use learned gated fusion rather than averaging or concatenation (ablation) | SIGNET · Phoenix14T | BLEU-4: 24.12 / 24.30 -> 27.82 | Improved |
| Use gated fusion rather than concatenation (ablation) | SIGNET · CSL-Daily | BLEU-4: 25.75 -> 28.51 | Improved |
| Choose source LLM initialisation for unseen-language datasets (ablation) | SIGNET · Phoenix14T | Test BLEU-4: 27.82 vs 27.89 vs 27.85 | No clear effect |
| Replace Geo-Sign with SIGNET | Geo-Sign · CSL-Daily | BLEU-4: 27.42 -> 28.51 | Improved |
Closest papers by meaning
- Scaling Sign Language Translation (NeurIPS 2024 · similarity 0.94)
- Lost in Translation, Found in Embeddings: Sign Language Translation and Alignment (NeurIPS 2026 · similarity 0.92)
- Sign2GPT: Leveraging Large Language Models for Gloss-Free Sign Language Translation (ICLR 2024 · similarity 0.92)
- LLMs are Good Sign Language Translators (CVPR 2024 · similarity 0.91)
- Bridging Sign and Spoken Languages: Pseudo Gloss Generation for Sign Language Translation (NeurIPS 2025 · similarity 0.91)
- Improving Gloss-free Sign Language Translation by Reducing Representation Density (NeurIPS 2024 · similarity 0.91)
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.
