Abstract
Syntactic query targeting exclusions are often imprecise. Extensive manual management of syntactic string variations is required. Consequently, relevant search queries are frequently over-filtered, and complex semantic variations are missed. Natural language matching guidelines are utilized to dynamically control search query targeting. Synthetic training data is generated by a large language model. This synthetic data includes simulated guidelines and labeled search queries. The generated data is utilized to train a low-latency serving model. Incoming search queries and contextual signals are converted into vector representations during online serving. The semantic intent of the incoming queries is evaluated against an internal representation of the matching guidelines by the trained serving model. Furthermore, complex natural language instructions are decoupled to instantiate deterministic geographic and brand controls, and to route specific creative messaging. The manual listing of syntactic query variations is eliminated, and strict online serving latency constraints are simultaneously satisfied.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Picozzi, Google Case Managers; Lennerz, Julia Maria; Giannoni, Federico; Tsvetkova, Desislava Mariova; van der Kamp, William Stephen; Lu, Wenjun; Szubert, Marcin Grzegorz; and Kiderle, Oliver Albert, "Natural Language Evaluation and Instantiation Controls for Search Targeting", Technical Disclosure Commons, (July 27, 2026)
https://www.tdcommons.org/dpubs_series/11151