ResearchPod Summary
Two-sided service marketplaces, which connect consumers with local professionals, have historically relied on rigid, deterministic request forms to capture provider preferences. As these platforms transition to AI-native probabilistic matching, they require a more flexible and explicit catalog of provider attributes. The authors address the challenge of reconstructing these implicit catalogs from legacy question-and-answer data, ensuring the new taxonomy is both interpretable to service providers and useful for marketplace decision-making.
The researchers implemented an automated, iterative system that treats each service occupation as an independent generation problem. This avoids the pitfalls of forcing diverse service domains into a single, global hierarchy. The system follows a propose-evaluate-keep loop:
This process is cross-family, meaning the models used for generation and evaluation differ from those used for critique and mutation, which helps prevent the reinforcement of shared biases.
A critical component of the system is the parity-mapping stage. Rather than performing literal translation, the system infers the underlying job attribute each legacy question was intended to measure. This allows the marketplace to map legacy data to the new taxonomy, providing a coverage signal and a human-reviewable artifact. The system has been deployed in production since April 2026, covering 132 occupations. By operating independently per occupation, the system surfaced specific catalog-hygiene issues—such as deprecated tags still being emitted by legacy systems—that would have been masked by a global, unified build.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.