DSIRM: Learning Query-Bridged Discrete Semantic Identifiers for E-commerce Relevance Modeling
Researchers have developed a new model called DSIRM to improve e-commerce search relevance by learning discrete semantic identifiers. This approach uses query-item interaction supervision to create relevance-aware item partitions and leverages generative LLMs to predict item identifiers from text. When deployed on Tmall's production data, DSIRM significantly improved offline AUC by 1.54% and showed positive online lifts in user click-through and conversion rates. AI
IMPACT Enhances e-commerce search relevance through learned discrete identifiers, potentially improving user experience and conversion rates.