Reframe generation
as a labeling problem.
The important modeling decision was to narrow the output space to the source name.
A generative API cleaned multilingual business names, but inference cost reached $70,000 per run and generated output could contain invented tokens. For this task, the cleaned name could be constructed by retaining or removing source tokens.
I trained a BERT token-labeling model to predict KEEP or DROP and reconstructed the cleaned name from the original text. That constraint directly addressed the risk of introducing new name tokens.
Cleaned name: Acme Café
Labels are hand-written to illustrate the design. This example does not send data to a model.
Multilingual examples, manual review, calibration, and batch inference on AWS Batch supported the production workflow.
Exact-match checks and error-slice review informed assessment. The reported outcome here is cost reduction; no numeric accuracy lift is claimed.
What this constraint solves—and where it needs care
Reconstructing from source text prevents invented name tokens. It does not, by itself, guarantee a correct name: the model can still remove useful text or retain clutter. Multilingual error review remains necessary. This formulation is best suited to cleanup tasks where the desired answer exists in the input, rather than tasks requiring correction or translation.
