Problem
A traveler often searches by a need, not by an attribute: «hotels with family-friendly beaches for summer vacation». A beach can be tied to a hotel through geography, but «family-friendly beaches» has no field in the catalog. A Boolean flag such as «free breakfast» also ignores what guests think about it. Labeling hotels by hand with every possible concept is too expensive, so search and recommendations cannot use such concepts.
Example from the article: a search such as «Luxury hotels in Geneva», or a personalized recommendation, needs hotels and concepts in one similarity space.
Solution in brief
Train a two-tower model: one tower embeds a hotel, the other embeds a concept. Supervision comes from guest reviews. Concepts mentioned in a hotel's review are positives, and concepts with negative mentions for that hotel are negatives, so no human labels are needed. At query time, look up the concept embedding and retrieve the nearest cached hotel embeddings.
Questions this case answers
- How to retrieve items for a free-text concept that has no structured attribute, using only the opinions users already wrote about the items, with no manual labels
- How to match a free-text travel concept such as «family-friendly beaches» to hotels when no structured attribute covers it?
- How to train a shared embedding space for items and concepts without paying for human annotation?
- How to serve concept-based hotel search cheaply at query time?
Tools and methods
- Two-tower model — separate hotel and concept encoders scored by inner product or cosine similarity
- Margin ranking loss — pulls a hotel toward a concept from its reviews and away from a negatively mentioned one
- Hard negative selection — keeps only negatives that violate the margin, so training is not trivial
- In-house review tagger — extracts concepts and their sentiment from review text
Google or a one-time email code. No password.