Research map / Large language models / Language, safety and interpretability
CiCo: Domain-Aware Sign Language Retrieval via Cross-Lingual Contrastive Learning
Yiting Cheng, Fangyun Wei, Jianmin Bao, Dong Chen, Wenqiang Zhang
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 |
|---|---|---|---|
| Use CiCo instead of SPOT-ALIGN's strongest SA-COMB baseline | Sign language retrieval · How2Sign test | T2V R@1: +22.4 | Improved |
| Use CiCo instead of SPOT-ALIGN's strongest SA-COMB baseline | Sign language retrieval · PHOENIX-2014T test | V2T R@1: +17.1 | Improved |
| Add the domain-aware/domain-agnostic sign encoder (ablation) | CiCo baseline · How2Sign test | T2V R@1: +1.8 | Improved |
| Add cross-lingual contrastive learning with fine-grained sign-to-word alignment (ablation) | CiCo with sign encoder · How2Sign test | T2V R@1: +20.7 | Improved |
| Add random-swap text augmentation (ablation) | CiCo without text augmentation · How2Sign test | T2V R@1: +2.6 | Improved |
| Apply random-delete text augmentation (ablation) | CiCo without text augmentation · How2Sign test | T2V R@1: 54.0 -> 52.7 | Hurt |
| Apply synonym-replacement text augmentation (ablation) | CiCo without text augmentation · How2Sign test | V2T R@1: 50.0 -> 47.5 | Hurt |
| Initialize vision and text Transformers from CLIP rather than randomly (ablation) | CiCo · How2Sign test | T2V R@1: +11.4 | Improved |
| Freeze shallow text Transformer layers instead of fine-tuning all layers (ablation) | CiCo · How2Sign test | T2V R@1: 56.6 -> 54.4 | Hurt |
| Use softmax mapping instead of mean or max for fine-grained sign-to-word alignment (ablation) | CiCo CLCL · How2Sign test | T2V R@1: 33.1 / 42.2 -> 56.6 | Improved |
| Combine domain-agnostic and domain-aware features with weighted sum rather than average (ablation) | CiCo sign encoder · How2Sign test | T2V R@1: 54.7 -> 56.6 | Improved |
| Use a longer maximum sign-clip feature sequence (ablation) | CiCo · How2Sign test | T2V R@1: 17.3 -> 38.2 -> 50.9 -> 53.6 -> 56.6 | Improved |
Closest papers by meaning
- Addressing Resource Scarcity across Sign Languages with Multilingual Pretraining and Unified-Vocabulary Datasets (NeurIPS 2022 · similarity 0.92)
- Lost in Translation, Found in Embeddings: Sign Language Translation and Alignment (NeurIPS 2026 · similarity 0.91)
- YouTube-SL-25: A Large-Scale, Open-Domain Multilingual Sign Language Parallel Corpus (ICLR 2025 · similarity 0.90)
- Uni-Sign: Toward Unified Sign Language Understanding at Scale (ICLR 2025 · similarity 0.89)
- Lost in Translation, Found in Context: Sign Language Translation with Contextual Cues (CVPR 2025 · similarity 0.89)
- Scaling Sign Language Translation (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.
