Reproduce “Semantic Retrieval for Product Search in E-Commerce” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Nikhil Kothari, Saksham Samdani, Ritam Mallick, Praveen Gupta, Ankit Vijay, Surender Kumar
Flipkart researchers developed an LLM-based semantic retrieval system for e-commerce product search, enhancing relevance by combining a dual-encoder architecture with a novel two-stage training pipeline and a comprehensive data strategy. The system improved offline MAP@8 by 5.57 points and NDCG@8 by 3.56 points, leading to a 2.39% increase in online Click-through Rate and a 2.62% rise in Orders.
Key insights
- A two-stage training pipeline, combining contrastive learning with a preference optimization objective like ROAR, effectively balances precision and efficiency for large-scale e-commerce semantic retrieval.
- Optimizing for graded relevance using an odds-ratio-based objective (ROAR) significantly enhances retrieval quality by learning fine-grained distinctions beyond binary relevance.
- A robust, multi-source data strategy, including LLM-driven augmentations and diverse real-world feedback, is critical for building scalable and performant e-commerce search systems, particularly for long-tail queries.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.