Siamese LLM Dual-Encoder with ROAR for Semantic Product Search in E-Commerce

Summary

This paper presents a Siamese LLM dual-encoder for semantic retrieval in e-commerce search, addressing challenges of short, noisy queries over large product catalogs. The model uses a two-stage training pipeline: contrastive learning with a false-negative margin mask to handle near-duplicate products, followed by Relative Odds Alignment for Retrieval (ROAR), a preference optimization objective that extends Bradley-Terry to variable-sized graded relevance groups. Training data progresses from substitute query-product pairs (coarse semantic supervision) to graded relevance annotations (fine-grained ranking). The system accurately retrieves exact matches while ordering substitutes and complementary products, with gains confirmed across query-frequency strata and business verticals, validated through live A/B deployment at scale.