Researchers have developed a novel method for generating high-quality training data for dense retrieval models, particularly for e-commerce sponsored search. This approach leverages disagreements between multiple retrieval systems to create structured training signals, including easy positives, hard positives, and hard negatives. The system combines multi-channel retrieval mining, a calibrated three-model cascade for graded relevance annotation, and a progressive curriculum training strategy using over 240 million examples. When deployed on Walmart's sponsored search, the trained BERT model demonstrated significant improvements, including a 5.1% increase in NDCG@10, a reduction in embarrassing retrievals, and positive impacts on ad spend, click-through rate, and conversion rates in an A/B test. AI
IMPACT This method offers a scalable blueprint for improving sponsored search relevance by replacing click-based training with LLM-annotated data.
RANK_REASON The cluster describes a research paper detailing a new method for training AI models. [lever_c_demoted from research: ic=1 ai=1.0]
Read on arXiv cs.IR (Information Retrieval) →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
