Research map / Vision-language and multimodal / Cross-modal learning
CVT-SLR: Contrastive Visual-Textual Transformation for Sign Language Recognition With Variational Alignment
Jiangbin Zheng, Yile Wang, Cheng Tan, Siyuan Li, Ge Wang, Jun Xia, Yidong Chen, Stan Z. Li
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 the pretrained textual module (ablation) | CVT-SLR with pretrained visual module · PHOENIX-2014 | WER: 20.4 -> 19.8; 20.7 -> 20.1 | Improved |
| Add the pretrained textual module (ablation) | CVT-SLR with pretrained visual module · PHOENIX-2014T | WER: +1.7%/+1.6% | Improved |
| Add the pretrained visual module (ablation) | CVT-SLR with pretrained textual module · PHOENIX-2014 | WER: 30.8 -> 19.8; 31.4 -> 20.1 | Improved |
| Add contrastive alignment loss (ablation) | CVT-SLR without pretrained VAE parameters · PHOENIX-2014 | WER: 21.1 -> 20.4; 21.4 -> 20.7 | Improved |
| Add pretrained VAE parameters (ablation) | CVT-SLR without contrastive alignment · PHOENIX-2014 | WER: 21.1 -> 20.2; 21.4 -> 20.3 | Improved |
| Add pretrained VAE parameters (ablation) | CVT-SLR without contrastive alignment · PHOENIX-2014T | WER: 21.8 -> 20.1; 22.0 -> 20.4 | Improved |
| Add contrastive alignment loss (ablation) | CVT-SLR with pretrained VAE parameters · PHOENIX-2014 | WER: 20.2 -> 19.8; 20.3 -> 20.1 | Improved |
| Add contrastive alignment loss (ablation) | CVT-SLR with pretrained VAE parameters · PHOENIX-2014T | WER: 20.1 -> 19.4; 20.4 -> 20.3 | Improved |
| Set alignment loss weight to 10.0 (ablation) | CVT-SLR, CTC weight 1.0 · PHOENIX-2014 | WER: 20.2 -> 19.8; 20.3 -> 20.1 | Improved |
| Increase alignment loss weight from 10.0 to 100.0 or 1000.0 (ablation) | CVT-SLR, CTC weight 1.0 · PHOENIX-2014 | WER: 19.8 -> 20.1 -> 20.5; 20.1 -> 20.6 -> 20.9 | Hurt |
| Use the full CVT-SLR method rather than the strongest single-cue baseline | PHOENIX-2014 · PHOENIX-2014 | WER: 20.8 -> 19.8; 21.0 -> 20.1 | Improved |
| Use the full CVT-SLR method rather than the strongest multi-cue baseline | PHOENIX-2014T · PHOENIX-2014T | WER: 20.2 -> 19.4; 20.4 -> 20.3 | Mixed |
Closest papers by meaning
- Conditional Variational Autoencoder for Sign Language Translation with Cross-Modal Alignment (AAAI 2024 · similarity 0.93)
- 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 0.89)
- Scaling Sign Language Translation (NeurIPS 2024 · similarity 0.89)
- LLMs are Good Sign Language Translators (CVPR 2024 · similarity 0.89)
- Improving Gloss-free Sign Language Translation by Reducing Representation Density (NeurIPS 2024 · 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.
