Problem
Airbnb guests look for specifics such as a workspace, reliable internet, a crib or a highchair. Many listings do not state these attributes, so what the host offers and what guests search for do not match. Asking hosts to fill in hundreds of possible attributes by hand is tedious. Airbnb also cut the impressions of Supplementary Review Flow (SRF) questions to protect the guest review experience, so guest-side data intake fell. Third-party vendors were another source, with their own limits.
Example from the article: guests who work remotely need to know whether a listing has a suitable workspace and reliable internet, but the listing page may not say so.
Solution in brief
Airbnb built the Listing Attribute Extraction Platform (LAEP), which reads anonymized guest text and works in three stages. A custom NER model finds entities. An embedding-similarity step maps each phrase to a standard attribute. A fine-tuned BERT classifier decides whether the attribute is present in the listing.
Questions this case answers
- How to build structured catalog attributes from free-text user content when manual entry and surveys cover too little
- How to map thousands of spelling variants of an entity onto a fixed taxonomy without labeling every phrase-attribute pair
- How to collect structured listing attributes when hosts do not fill them in and guest questionnaires get little response?
- How to extract amenities and other entities from free guest text and map them to a fixed attribute taxonomy?
Tools and methods
- Named Entity Recognition (CNN) — finds amenity, facility, hospitality, location and structure phrases in free text
- Word2Vec fine-tuned on Airbnb text — embeds attributes and phrases for mapping without labeled pairs
- Cosine similarity with a confidence threshold — picks the closest attribute or returns "No Mapping"
- BERT — classifies attribute presence (YES, Unknown, NO) from local context
Google or a one-time email code. No password.