This paper proposes BroadGen, a novel framework for enhancing the efficiency and effectiveness of keyword recommendations in sponsored search advertising. To address the challenges of existing exact match keyword methods, such as high management costs, limited targeting scope, and evolving search query patterns, and the challenges of broad match keyword methods, such as low targeting accuracy and insufficient supervision signals, BroadGen utilizes historical search query data to recommend efficient and effective broad match keywords. In particular, it demonstrates exceptional ability to maintain search term stability over time through token-correspondence modeling, demonstrating its ability to serve millions of sellers and over 2.5 billion products daily on eBay.
