Research map / Label-efficient and robust learning / Recognition and long-tail
MSA2: Multi-task Framework with Structure-aware and Style-adaptive Character Representation for Open-set Chinese Text Recognition
Yangfu Li, Hongjian Zhan, Qi Liu, Li Sun, Yu-Jie Xiong, Yue Lu
This paper tried a common change: Add a contrastive loss
add glyph-centric contrastive learning iterations to SAGE · BCTR line accuracy · line accuracy (%) Mixed
What 1,317 accepted papers report for this change
- 1,253 improved
- 22 no clear effect
- 16 mixed
- 26 hurt
- Beyond Looking Up, Try Looking Around: Harmonizing Global Structure and Local Consistency in Optimal Transport for Short Text Clustering ICML 2026 · add contrastive learning · SearchSnippets · ACC: 56.65 -> 80.12Improved
- SleepLM: Natural-Language Intelligence for Human Sleep ICML 2026 · add contrastive objective to captioning (CoCa vs Cap) · sleep event IoU · IoU: 29.2 -> 27.4Hurt
- LIMSSR: LLM-Driven Sequence-to-Score Reasoning under Training-Time Incomplete Multimodal Observations ICML 2026 · add contrastive loss L_con · FS1000 average · Spearman correlation: 0.759 -> 0.789Improved
See this change in detailMatched by the wording of each paper's own result, read by AI from the full text. Papers report gains more often than failures, so this shows what gets published, not the chance a change works.
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 glyph task to linguistics-only recognition (ablation) | MSA2 · Compound characters, standard forms | recall (%): 91.24 -> 91.95 | Improved |
| add glyph task to linguistics-only recognition (ablation) | MSA2 · Compound characters, other forms (Scene/Handw) | recall (%): 88.38 -> 89.82 | Improved |
| add linguistics task to glyph-only recognition (ablation) | MSA2 · Compound characters, other forms (Scene/Handw) | recall (%): 87.93 -> 89.82 | Improved |
| add glyph task to linguistics-only recognition (ablation) | MSA2 · Basic characters, standard forms | recall (%): 92.11 -> 93.43 | Improved |
| add linguistics task to glyph-only recognition (ablation) | MSA2 · Basic characters, other forms | recall (%): 89.26 -> 90.41 | Improved |
| replace one-hot labels with MSA2 (ablation) | one-hot baseline · BCTR closed-set, HierCode config average | accuracy (%): 70.0 -> 75.4 | Improved |
| replace HierCode with MSA2 under identical configuration (ablation) | HierCode · BCTR closed-set, HierCode config average | accuracy (%): 74.2 -> 75.4 | Improved |
| replace HierCode with MSA2 on handwritten open-set characters (ablation) | HierCode · ICDAR-char open-set (HWDB, m=1000) | character accuracy (%): 20.71 -> 26.13 | Improved |
| replace HierCode with MSA2 on non-Chinese character recall (ablation) | HierCode · BCTR/ICDAR non-Chinese recall (Handw. average) | recall (%): 86.61 -> 89.22 | Improved |
| replace HierCode with MSA2 on Chinese character recall (ablation) | HierCode · BCTR/ICDAR Chinese recall (Handw. average) | recall (%): 92.35 -> 94.38 | Improved |
| use SACE dynamic structure code instead of random structure code (ablation) | SACE · ICDAR-char compound characters | accuracy rate (%): 94.03 -> 97.35 | Improved |
| add glyph-centric contrastive learning iterations to SAGE (ablation) | SAGE · BCTR line accuracy | line accuracy (%) | Mixed |
Closest papers by meaning
- Chinese Text Recognition with A Pre-Trained CLIP-Like Model Through Image-IDS Aligning (ICCV 2023 · similarity 0.88)
- GlyphMastero: A Glyph Encoder for High-Fidelity Scene Text Editing (CVPR 2025 · similarity 0.86)
- An Exemplar-based Framework for Chinese Text Recognition (AAAI 2025 · similarity 0.86)
- OCRBench v2: An Improved Benchmark for Evaluating Large Multimodal Models on Visual Text Localization and Reasoning (NeurIPS 2025 · similarity 0.85)
- Linguistics-aware Masked Image Modeling for Self-supervised Scene Text Recognition (CVPR 2025 · similarity 0.84)
- ASCIIEval: Benchmarking Models' Visual Perception in Text Strings via ASCII Art (ICLR 2026 · similarity 0.84)
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.
