Research map / Large language models / Language, safety and interpretability
Visual Alignment Pre-training for Sign Language Translation
Peiqi Jiao, Yuecong Min, Xilin Chen
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 |
|---|---|---|---|
| add VAP visual alignment pre-training to skeleton-based SLT baseline | CoSign-1s baseline · PHOENIX14T Test | BLEU-4: 19.72 -> 26.16 | Improved |
| add VAP to baseline | CoSign-1s baseline · OpenASL Test | BLEU-4: 10.52 -> 21.23 | Improved |
| re-initialize the translation module of baseline once (ablation) | gloss-free SLT baseline · PHOENIX14T | BLEU-4: 19.72 -> 23.04 (Test) | Improved |
| re-initialize translation module more times (iter 3) (ablation) | gloss-free SLT baseline · PHOENIX14T Dev | BLEU-4: 23.75 -> 22.61 | Mixed |
| pre-train visual encoder with L_Align (ablation) | gloss-free SLT baseline · PHOENIX14T Test | BLEU-4: 19.72 -> 22.94 | Improved |
| add L_SLT to L_Align pre-training (ablation) | VAP pre-training · PHOENIX14T Test | BLEU-4: 22.94 -> 24.62 | Improved |
| fine-tune with mBART pre-trained translation module instead of from scratch (ablation) | VAP · PHOENIX14T Dev | BLEU-4: 24.38 -> 26.23 | Improved |
| add pseudo-gloss-to-text (P2T) fine-tuning of mBART (ablation) | VAP · PHOENIX14T Dev | BLEU-4: 26.23 -> 26.48 | No clear effect |
| keep L_Align during end-to-end fine-tuning (ablation) | VAP · PHOENIX14T Dev | BLEU-4: 26.48 -> 26.62 | No clear effect |
| use max pooling instead of average pooling for similarity aggregation (ablation) | VAP · PHOENIX14T Test | BLEU-4: 17.13 -> 25.54 | Improved |
| replace pre-trained embedding layer with Transformer text encoder (ablation) | VAP text encoder · PHOENIX14T Dev | BLEU-4: 26.62 -> 25.52 | Hurt |
| use pseudo glosses with embedding-initialized CTC in place of real glosses (ablation) | SLT trained from scratch · PHOENIX14T Test | BLEU-4: 26.14 vs 27.00 with gloss | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Sign2GPT — Sign2GPT: Leveraging Large Language Models for Gloss-Free Sign Language Translation (ICLR 2024) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Lost in Translation, Found in Embeddings: Sign Language Translation and Alignment (NeurIPS 2026 · similarity 0.91)
- Bridging Sign and Spoken Languages: Pseudo Gloss Generation for Sign Language Translation (NeurIPS 2025 · similarity 0.91)
- LLMs are Good Sign Language Translators (CVPR 2024 · similarity 0.91)
- Conditional Variational Autoencoder for Sign Language Translation with Cross-Modal Alignment (AAAI 2024 · similarity 0.91)
- Sign2GPT: Leveraging Large Language Models for Gloss-Free Sign Language Translation (ICLR 2024 · similarity 0.90)
- Scaling Sign Language Translation (NeurIPS 2024 · similarity 0.90)
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.
