Semantic Retrieval for Product Search in E-Commerce
Researchers have developed a new semantic retrieval system for e-commerce product search, designed to handle imprecise user queries and large product catalogs. The system utilizes a Siamese LLM dual-encoder trained in a two-stage process, beginning with contrastive learning and progressing to a preference optimization objective called Relative Odds Alignment for Retrieval (ROAR). This approach aims to accurately retrieve exact matches while effectively ranking substitute and complementary products, with demonstrated success through large-scale A/B testing. AI
IMPACT Enhances product search capabilities by improving relevance and ranking for e-commerce platforms.