![[2602.17654] Mine and Refine: Optimizing Graded Relevance in E-commerce Search Retrieval](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.17654] Mine and Refine: Optimizing Graded Relevance in E-commerce Search Retrieval
Summary
The paper presents a two-stage framework called 'Mine and Refine' for optimizing graded relevance in e-commerce search retrieval, enhancing the effectiveness of semantic text embeddings.
Why It Matters
As e-commerce continues to grow, improving search retrieval systems is crucial for user engagement and business success. This research addresses the challenge of graded relevance, where users accept various degrees of match quality, thus providing a scalable solution for better search outcomes.
Key Takeaways
- Introduces a two-stage contrastive training framework for e-commerce search.
- Focuses on graded relevance, allowing for substitutes and complements in search results.
- Implements a multilingual Siamese two-tower retriever for enhanced semantic understanding.
- Utilizes engagement-driven auditing to reduce noise in training data.
- Demonstrates statistically significant improvements in retrieval relevance and business impact through A/B testing.
Computer Science > Information Retrieval arXiv:2602.17654 (cs) [Submitted on 19 Feb 2026] Title:Mine and Refine: Optimizing Graded Relevance in E-commerce Search Retrieval Authors:Jiaqi Xi, Raghav Saboo, Luming Chen, Martin Wang, Sudeep Das View a PDF of the paper titled Mine and Refine: Optimizing Graded Relevance in E-commerce Search Retrieval, by Jiaqi Xi and 4 other authors View PDF HTML (experimental) Abstract:We propose a two-stage "Mine and Refine" contrastive training framework for semantic text embeddings to enhance multi-category e-commerce search retrieval. Large scale e-commerce search demands embeddings that generalize to long tail, noisy queries while adhering to scalable supervision compatible with product and policy constraints. A practical challenge is that relevance is often graded: users accept substitutes or complements beyond exact matches, and production systems benefit from clear separation of similarity scores across these relevance strata for stable hybrid blending and thresholding. To obtain scalable policy consistent supervision, we fine-tune a lightweight LLM on human annotations under a three-level relevance guideline and further reduce residual noise via engagement driven auditing. In Stage 1, we train a multilingual Siamese two-tower retriever with a label aware supervised contrastive objective that shapes a robust global semantic space. In Stage 2, we mine hard samples via ANN and re-annotate them with the policy aligned LLM, and introduce a...
