Helping Home Shoppers Find a Home to Love Through Home Insights

Zillow8 Aug 2022

Problem

On a real-estate marketplace, shoppers still browse dozens or hundreds of listings after search has narrowed the list. Most descriptions look alike, and in zip codes with high inventory it is hard to remember what made Home A different from Home B. Search ranking has already run, yet the user still has to compare descriptions one by one.

Example from the article: a listing in Florida is more likely to get ocean-related or beach-related tags than a listing in Arizona, because the tags follow what people search for in each region.

Solution in brief

Zillow's Home Insights pipeline pulls key phrases from listing descriptions with a fine-tuned BERT named entity recognition (NER) model. It then drops phrases that are unclear without context and ranks the rest by similarity to keywords of high-traffic search landing pages in the listing's region. The top tags appear on listing photocards and detail pages.

Questions this case answers
  • How to surface the few distinctive attributes of each item in a catalog of near-identical free-text descriptions, ordered by what users in the same segment currently search for
  • How to turn free-text item descriptions into short, standalone, display-ready tags without hand-built taxonomies
  • How to help shoppers tell similar listings apart when browsing hundreds of near-identical item descriptions?
  • How to extract short, self-explanatory feature tags from free-text item descriptions?

Tools and methods

  • BERT fine-tuned for named entity recognition (NER) — marks the phrases in a description that count as "insights"
  • Part-of-speech tagging and a blocklist of undesirable terms — remove malformed or context-dependent phrases
  • Ranking model on search traffic signals — selects the top tags per listing using landing pages and keyword searches by region
Continue reading — it's free

Google or a one-time email code. No password.