Data‑Driven Synonym Transformation for Keyword Matching in Search Advertising

Search advertising involves three parties—users, advertisers, and the search engine—where advertisers submit bid keywords (often called queries) and users issue search queries. The engine matches queries to keywords according to advertiser‑specified matching modes (exact, phrase, or smart), then ranks ads based on bid, quality, and CTR.

The core problem, keyword matching, requires returning all keywords that satisfy a given matching mode for a query. This task faces three major challenges: a semantic gap between user and advertiser expressions, determining the correct matching mode, and ensuring engineering scalability for massive query‑keyword volumes.

Synonym transformation is introduced as a key solution, with three primary application scenarios: (1) synonym matching to expand query coverage, (2) keyword‑side queue compression to reduce online verification load, and (3) query‑side normalization and rewriting to improve hit rates in high‑frequency term tables.

Three‑step data‑driven workflow for synonym transformation:

Collect large synonym data sources from external logs (click, session, collaborative filtering) and internal commercial logs.

Train models—options include sequence‑to‑sequence (S2S) translation models, dual‑tower semantic similarity models, or graph‑based link‑prediction models—using weakly supervised and a small amount of manually labeled data.

Use the trained model to generalize and recall additional synonym variants, applying post‑processing such as synonym reduction (normalization) and expansion to control redundancy and latency. Efficiency techniques include synonym reduction by normalizing both query and keyword spaces (removing redundant tokens, applying concept‑level mapping) and synonym expansion via concept templates or nearest‑neighbor search in a learned semantic space. Robustness is enhanced through adversarial training: generating adversarial examples with GAN‑style perturbations and mixing them into training data to improve model stability against small input variations. The article concludes that a data‑driven approach—leveraging massive weak supervision, multi‑stage pre‑training (e.g., ERNIE Large), and fine‑tuning—significantly outperforms feature‑driven shallow models, while acknowledging open problems such as handling the vast unseen query‑keyword space and further improving model robustness.