Research map / Label-efficient and robust learning / Semi-supervised learning
Semi-Supervised Synthetic Data Generation with Fine-Grained Relevance Control for Short Video Search Relevance Modeling
Haoran Li, Zhiming Su, Junyan Yao, Enwei Zhang, Yang Ji, Yan Chen, Kan Zhou, Chao Feng, Jiao Ran
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 SSRA synthetic data to annotated training data | Qwen3-Embedding-4B · Short video retrieval | nDCG@10: 73.20 -> 74.47 | Improved |
| add prompt-based SyCL Modified synthetic data | Qwen3-Embedding-4B · Short video pair classification | AP average: 69.57 -> 67.95 | Hurt |
| add vanilla SFT query-model synthetic data | Qwen3-Embedding-4B · Short video pair classification | AP average: 69.57 -> 70.87 | Improved |
| add SSRA over vanilla SFT synthesis | Qwen3-Embedding-4B · Short video pair classification | AP average: 70.87 -> 71.52 | Improved |
| add SSRA stage 1 (score-based re-annotation) (ablation) | SSRA · Short video retrieval | nDCG@10: 74.02 -> 74.23 (4B) | Improved |
| add SSRA stage 2 (relevance alignment filtering) (ablation) | SSRA · Short video pair classification | AP: 71.30 -> 74.92 (4B) | Improved |
| add SSRA stage 1 only (without stage 2) for pair classification (ablation) | Qwen3-Embedding-0.6B · Short video pair classification | AP: 70.70 -> 70.77 | No clear effect |
| use stage 1 re-labeled data for query model (ablation) | query model · 10,000 sampled documents | duplicate query rate: 6.57% -> 5.20% | Improved |
| add stage 2 training to query model (ablation) | query model · 200 documents, human annotated | queries matching intended label (label 1): 81 / 200 -> 130 / 200 | Improved |
| add stage 2 training to query model (ablation) | query model · 200 documents, human annotated | queries matching intended label (label 3): 189 / 200 -> 178 / 200 | Hurt |
| include intermediate relevance labels 1 and 2 in synthetic data (multi-level vs binary) (ablation) | Qwen3-Embedding-4B · Short video retrieval | nDCG@10: 73.73 -> 74.23 | Improved |
| include intermediate relevance labels 1 and 2 in synthetic data (multi-level vs binary) (ablation) | Qwen3-Embedding-4B · Balanced pair classification | AP@>=3: 58.39 -> 55.25 | Hurt |
Closest papers by meaning
- Mitigating Semantic Collapse in Partially Relevant Video Retrieval (NeurIPS 2025 · similarity 0.86)
- Generative Video Diffusion for Unseen Novel Semantic Video Moment Retrieval (AAAI 2025 · similarity 0.85)
- ReSpec: Relevance and Specificity Grounded Online Filtering for Learning on Video-Text Data Streams (CVPR 2025 · similarity 0.85)
- NUDGE: Lightweight Non-Parametric Fine-Tuning of Embeddings for Retrieval (ICLR 2025 · similarity 0.84)
- Needle In A Video Haystack: A Scalable Synthetic Evaluator for Video MLLMs (ICLR 2025 · similarity 0.84)
- TempMe: Video Temporal Token Merging for Efficient Text-Video Retrieval (ICLR 2025 · 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.
