Research map / Vision-language and multimodal / Human and embodied modeling
UNISON: Unpaired Cross-Lingual Image Captioning
Jiahui Gao, Yi Zhou, Philip L. H. Yu, Shafiq Joty, Jiuxiang Gu
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 UNISON instead of Graph-Align followed by Google translation | Chinese image captioning without a caption corpus · COCO-CN test | BLEU-1: 39.2 -> 44.9 | Improved |
| Use UNISON instead of FC-2k trained on pseudo Chinese captions | Chinese image captioning with a caption corpus · COCO-CN test | BLEU-1: 60.4 -> 63.4 | Improved |
| Map scene-graph nodes in encoding space rather than symbolically translate nodes (ablation) | Cross-lingual graph mapping · MT corpus test | BLEU-1: 26.6 -> 28.1 | Improved |
| Add subgraph-level contextual mapping to word-level mapping (ablation) | GM_word · MT corpus test | BLEU-1: 28.1 -> 29.2 | Improved |
| Add full-graph mapping to word- and subgraph-level mapping (ablation) | GM_word+sub · MT corpus test | BLEU-1: 29.2 -> 29.6 | Improved |
| Replace concatenation and an FC layer with self-gated adaptive fusion (ablation) | HGM_base · MT corpus test | BLEU-1: 29.6 -> 30.4 | Improved |
| Add KL joint training to align source- and target-language scene-graph features (ablation) | HGM · MT corpus test | BLEU-1: -0.5 | Improved |
| Add cross-modal feature mapping (ablation) | HGM · COCO-CN test | BLEU-1: 41.1 -> 44.9 | Improved |
| Add cross-modal feature mapping (ablation) | GM_word · COCO-CN test | BLEU-1: 40.1 -> 43.1 | Improved |
| Add cross-modal feature mapping (ablation) | GM_word+sub · COCO-CN test | BLEU-1: 37.3 -> 40.6 | Improved |
| Add cross-modal feature mapping (ablation) | HGM_base · COCO-CN test | BLEU-1: 38.0 -> 39.8 | Improved |
| Use contextual graph-level mapping without adaptive self-gating on noisy image scene graphs (ablation) | GM_word+sub and HGM_base · COCO-CN test | BLEU-1: GM_word 40.1; GM_word+sub 37.3; HGM_base 38.0 | Hurt |
Closest papers by meaning
- Unpaired Image Captioning by Language Pivoting (ECCV 2018 · similarity 0.91)
- Dynamic Adapter with Semantics Disentangling for Cross-lingual Cross-modal Retrieval (AAAI 2025 · similarity 0.88)
- Semantic and Expressive Variations in Image Captions Across Languages (CVPR 2025 · similarity 0.88)
- CL2CM: Improving Cross-Lingual Cross-Modal Retrieval via Cross-Lingual Knowledge Transfer (AAAI 2024 · similarity 0.88)
- Improving Cross-Modal Alignment with Synthetic Pairs for Text-Only Image Captioning (AAAI 2024 · similarity 0.87)
- LLM4SGG: Large Language Models for Weakly Supervised Scene Graph Generation (CVPR 2024 · similarity 0.86)
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.
