Query Expansion for Search

2 posts · May 2022 – Nov 2022

What this trail covers

Grab has developed a query expansion framework to improve search relevance and user experience. This framework expands user queries with semantically related keywords derived from manual annotations of top keywords and data mining of user rewrite behavior. The expanded queries are then used in a multi-recall and result fusion search architecture. Continuous monitoring and a feedback loop ensure the relevance and up-to-dateness of expansion pairs. Future work includes enhancing data mining with more advanced techniques.

All Grab trails

Nov 2022Query expansion based on user behaviourThis post details the implementation of a query expansion framework for Grab's search system. It describes two methods for generating expansion candidates: manual annotation of top keywords and data mining based on user rewrites. The data mining approach leverages user search session data to identify effective rewrites based on heuristics like adjacent queries, click behavior, time intervals, and frequency. The post also outlines the expansion matching architecture within the existing search system and the continuous monitoring and feedback loop for maintaining relevance. Results from A/B experiments show an uplift in CTR and conversion rate, with the data mining approach performing better.
May 2022Search architecture revampThis post details the revamp of Grab's search architecture from a legacy text-matching system to a multi-layer approach including boosting, multi-recall, and ranking. Key improvements include intent recognition, context consideration, effective blending of results from multiple business verticals, and a modularized plugin-based implementation to enable quick iteration and reuse of components for new search use cases. The new architecture supports embedding-based recall and a mixed Ads/organic framework.
Advertisement