Reproduce “CORE: A Unified Cascaded Ordinal Relevance Estimation Framework for E-commerce 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
Zhi Jin, Xi Wang, Yunfei Li, Guojun Liu, Qingsong Hua, Wei Lin
Meituan researchers developed CORE, a cascaded ordinal relevance estimation framework that reformulates e-commerce search relevance as sequential binary classifications. This approach improved offline accuracy by 1.17 percentage points for BERT models and led to a 0.20% relative increase in NDCG@5 and a 15.9% relative reduction in severe online bad cases.
Key insights
- Re-framing ordinal relevance as cascaded binary classifications better aligns models with the inherent semantic hierarchy and asymmetric decision boundaries of relevance levels.
- Employing step-level reinforcement learning allows for fine-grained credit assignment, leading to improved training stability and higher accuracy in LLM-based structured reasoning.
- Advanced reasoning capabilities from large language models can be effectively distilled into more efficient BERT-based models for practical online deployment through PostCoT.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.