Automatic induction of world-grounded SLS recipes

Develop a method to automatically induce the world-grounded factorization of each verified QA unit’s representative question into linguistic and semantic slots for Staged Linguistic Seeding, thereby replacing human operator authoring while preserving the bridge between curated questions and real-user query formulations.

Background

Staged Linguistic Seeding (SLS) currently depends on a human operator who factorizes each verified unit’s representative question into world-grounded slots and provides the resulting recipe to gpt-4.1-mini, which generates candidate query variants that are subsequently filtered by a human gate. This human-authored grounding is presented as the mechanism that captures reformulations whose surface words differ substantially from those in the canonical question.

The paper identifies automatic induction of this factorization as the principal unresolved issue. Solving it would reduce the manual authoring burden and make the verified-unit QA pipeline more scalable to substantially larger FAQ collections, while retaining the semantic coverage that distinguishes SLS from automatic doc2query-style expansion.