Reproduce “CroPS: Improving Dense Retrieval with Cross-Perspective Positive Samples in Short-Video Search” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ao Xie, Jiahui Chen, Quanzhi Zhu, Xiaoze Jiang, Zhiheng Qin, Enyun Yu, Han Li
Kuaishou Technology researchers developed CroPS, a dense retrieval framework that addresses the filter bubble effect in short-video search by augmenting positive samples from diverse, cross-perspective sources including LLM-generated world knowledge. This approach, which includes hierarchical label assignment and an H-InfoNCE loss, improved offline recall by 9.3% and, in live A/B tests, increased Click-Through Rate by 0.869% and reduced query reformulation by 0.646%.
Key insights
- Enriching positive samples from diverse, cross-perspective sources effectively breaks the filter bubble without complex architectural changes.
- Fine-grained hierarchical relevance labeling is crucial for training robust retrieval models capable of understanding nuanced semantic relationships beyond binary relevance.
- Large Language Models can serve as powerful 'pseudo-retrievers' to synthesize external world knowledge, significantly augmenting internal platform data and improving semantic coverage.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.