Skip to content
conftrace_
2026 EMNLP EMNLP 2026

From Proprietary to Public: Dataset Declassification for Question Answering

AbstractTraining Question Answering (QA) systems on real customer interaction data is often infeasible to reproduce or share due to privacy and intellectual property constraints. We introduce dataset declassification, a framework that replaces proprietary QA datasets with public, semantically equivalent counterparts that preserve training and evaluation utility while satisfying typical legal and privacy requirements. The approach maps proprietary questions to public equivalents via large-scale retrieval and reconstructs answers from domain-matched corpora using LLM-based annotation. Experiments on four QA benchmarks show that models trained and evaluated on declassified datasets achieve performance within 𝛥 ≈ 0 of the original data. The method targets datasets whose underlying information needs already exist in public knowledge sources (e.g., knowledge available on the web and used by public-facing services). We further provide explicit applicability conditions showing that declassification succeeds when sufficient semantic overlap exists between proprietary and public corpora, and degrades predictably when it does not.

More from EMNLP 2026

All EMNLP 2026 papers →