Problem
Yandex Maps holds 40 million organizations, 200 million reviews and 250 million photos. Its recommendation feed shows places with a generated caption and a photo, mostly for discovery. Nothing tied a place to a season or holiday. The team had to find what matters in a given month, and the first caption and photo candidates were often dull, silly or off-season.
Example from the article: a cafe and the palace of Tsar Alexei Mikhailovich showed up in the "May holidays" filter, and users could not see why. The reasons were a summer veranda and an apple orchard.
Solution in brief
Extract aspects from reviews with two LLM steps and cluster them. Rank aspects by their share of mentions in a month versus the year to find seasonal ones. For each seasonal aspect, generate captions from reviews, filter out boring, silly and 18+ ones, and rank the rest. Pick photos by image-text similarity.
Questions this case answers
- How to find which aspects of an item matter in a given season from user-written texts and feed seasonal items into a recommendation feed
- How to filter generated short captions for boring, nonsensical and inappropriate outputs before ranking them
- How to find which aspects of a place matter in a given season or month from user reviews
- How to generate short, non-boring recommendation captions for places from reviews with an LLM
Tools and methods
- Aspect extraction and sentiment (YaGPT, p-tune) — two models over each review
- TF-IDF over aspects within a category — finds what is distinctive for an organization
- Month-versus-year share of aspect mentions — surfaces seasonal aspects
- Few-shot YaGPT ranking — picks the best candidate caption
- LLM stupidity classifier and CatBoost boringness classifier — filters weak captions
- Image-text embeddings and cosine similarity — filters and ranks photos
Google or a one-time email code. No password.
