Problem
Grab's food search results sometimes included merchants that were not yet operational or were outside the delivery radius, and sometimes offered no alternatives at all. The system also had trouble with typos, keywords in different languages, synonyms and word spacing. Grab calls the outcome a suboptimal user experience.
Example from the article: The query understanding module expands "KFC" to "fried chicken", and both keywords are searched.
Solution in brief
Expand each query to at most three related keywords and send the original and the expansions to the engine together. The expansion dictionary comes from two sources: manual annotation of the top frequent keywords and pairs mined from users' own query rewrites in search sessions. The mined pairs beat manual annotation in an A/B test in 6 countries.
Questions this case answers
- How to expand a search query with related queries when lexical similarity cannot capture the relation between what users type and what they want
- How to mine query-to-query relations from session logs without manual labeling
- How to recall more relevant results for a search query when the user's words do not match the catalog (typos, other language, synonyms, a cuisine instead of a merchant)?
- How to build a query expansion dictionary from search logs without labeling every query by hand?
Tools and methods
- User rewrite mining — turns adjacent query pairs in sessions into expansion candidates with simple filters
- Manual annotation of top keywords — covers the fat head of traffic with curated expansions
- Multi-recall with result fusion — searches the original query and each expansion separately, then merges results
- A/B test — control, manual-only and mined expansions compared on click-through and conversion
Google or a one-time email code. No password.