Problem
Users of a map service read only the first two or three reviews, sort negative ones to the top, read a couple more and decide. So 25 thousand reviews on one large Moscow shopping mall go unread, and the visitor gets an incomplete, sometimes misleading picture. The team wants the system to read reviews for the user and show, per organization, which aspects (food, staff, location) appear in how many reviews and how satisfied people are with each.
Example from the article: the first 8 aspects, for cafes and restaurants only, covered 5% of organization card impressions.
Solution in brief
One BERT model takes the aspect name plus the review text and returns two scores: whether the aspect is mentioned and the sentiment toward it. A self-attention row of the aspect token shows which words to highlight as evidence. New aspects are added through an iterative pipeline: candidate aspects from frequent words and bigrams, active-learning selection of reviews, crowd annotation, fine-tuning, quality check.
Questions this case answers
- How to summarize long user-generated texts into per-attribute counts and sentiment when readers skim only a few items
- How to grow a text-attribute classifier from a handful of attributes to hundreds with limited annotation budget and confusable neighbouring concepts
- How to show users what reviews say about a place when they read only the first few reviews?
- How to extract aspects (food, staff, location) and the sentiment toward each aspect from review text?
Tools and methods
- BERT — fine-tuned on "aspect name + review text" to predict mention and sentiment.
- Self-attention heads — vote over one encoder layer to locate the evidence token.
- Active learning — send for labeling only reviews in the model's gray zone.
- Toloka — crowd annotation of "is the aspect mentioned" with category and aspect description.
- Remorph — Yandex regex-over-words library, the rejected baseline.
Google or a one-time email code. No password.
