Task-Adapted Retrieval Boosts Geographic Entity Search Accuracy

People search applications face a significant challenge in accurately mapping diverse, free-form location phrases provided by users to standardized geographic entities used for structured retrieval. Existing lexical standardizers often fail when encountering aliases, misspellings, metropolitan area expressions, or ambiguous place names. This research frames the problem as a graded, set-valued entity retrieval task over a fixed ontology, identifying three critical design requirements: distinguishing between identity-preserving variations and knowledge-dependent aliases, managing false negatives among valid same-name entities, and separating stable transformations from mutable entity knowledge. To address these challenges, the researchers developed a prompt-asymmetric bi-encoder model. This model incorporates calibrated alias support, uses bounded ambiguity-aware negative examples during training, and features editable entity documents that allow for localized updates without requiring a full model retraining. Evaluated on both a production-derived benchmark and a public GeoNames transfer task, the task-adapted model showed substantial improvements over frozen encoders and standard token baselines. Specifically, it significantly raised relevant Precision@1 in a blinded human comparison on a production challenge set, particularly for non-canonical queries, while maintaining performance on frequent queries. These results indicate that task-adapted geographic entity retrieval is a practical and effective replacement for traditional taxonomy-based standardizers.