Problem
A marketplace receives about 100 thousand merchant offers a day, and a content team turns them into product cards by hand. The content department measured 17 minutes per card in 80% of cases, which would take about 3,500 people per 8-hour day. Offers wait in a queue, stock runs out meanwhile, and duplicates slip through. The work has three manual steps: choosing a category (5 minutes), matching merchant attributes to site attributes (10 minutes), and checking the offer group for duplicates (2 minutes).
Example from the article: on SberMegaMarket (35 million products on the storefront), three assisting models cut card creation time by more than 40% (45% by precise count).
Solution in brief
Each step gets its own model that assists the content manager. A typifier ranks categories as a retrieval task, an attribute matcher is a cross-encoder, and a grouper clusters offers. Groups already checked by people are frozen, so new model versions do not move what people approved.
Questions this case answers
- How to assign a record to one of thousands of categories in a tree that changes weekly by retrieving and re-scoring candidates instead of training a classifier
- How to group duplicate records of the same item into stable group IDs when groups already approved by people must not move after model updates
- How to speed up product card creation on a marketplace with ML
- How to assign a merchant offer to a category in a tree of thousands of categories that change weekly
Tools and methods
- Retrieval plus re-scoring cascade — vector search for top-M candidates, re-scoring, statistical features, final ranker; used for categories and for offer-to-delegate matching
- Cross-encoder — matches attributes to merchant offer text
- Connectivity graph with thresholds — connected components become offer groups
- Delegates — one representative per group for daily regrouping
- Yandex.Toloka — online quality evaluation
- F1 and share of fully matched groups — grouper metrics
Google or a one-time email code. No password.
