ReCoPII: A Relation-Grounded Framework for Building Contextual PII Masking Datasets

arXiv:2610.35925 [cs.CL] · Submitted 14 October 2026 · Taehwan Lee

Subjects: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)

Overview

Training contextual masking systems for personally identifiable information (PII) requires examples that distinguish maskable from preservable mentions of the same entity type. Privacy restrictions, however, limit access to real data, making synthetic dataset construction a practical necessity in 2026 as regulatory scrutiny of training data intensifies.

The Limitation of Prior Augmentation

Author-based augmentation (SPY) labels mentions unrelated to the author as non-PII. This approach presumes a single data subject and overlooks other individuals whose mentions may still require masking.

The ReCoPII Framework

We propose ReCoPII (Relation-grounded Contextual PII masking), a framework that builds such data without this single-subject premise. ReCoPII:

Plans ontology-constrained entity relations.
Writes insertion sentences with placeholders.
Reconstructs annotations with code.
Checks the result with a separate LLM.
Derives MASK/KEEP labels from entity relations.

ReCoPII-Ko Corpus and Validation

Applied to Korean documents from six domains, the framework yields ReCoPII-Ko, a corpus with relation, privacy, and MASK/KEEP annotations, which we validate.

Across 300 pairs, four annotators agree with 88.42% of its sampled KEEP labels, compared with 26.27% for SPY-style augmentation of the same sources.

Experimental Results

Classifiers trained on ReCoPII over-mask less than those trained under the SPY-style reference on a matched test.
Contextual masking fine-tuned jointly with a PII detector removes most over-masking of detection alone within a fixed leakage budget.
On human-verified entity records, the four detectors meeting that budget over-mask at most 5.04% with at most 4.18% leakage.
The detector is comparable to XLM-R-large on KDPII with about a quarter of its parameters.

Metadata

Cite as: arXiv:2610.35925 [cs.CL]
DOI: https://doi.org/10.48550/arXiv.2610.35925 (arXiv-issued DOI via DataCite, pending registration)
Submission history: [v1] Wed, 14 Oct 2026 09:48:32 UTC (6,601 KB)

via ArXiv CL+LG

ReCoPII: A Relation-Grounded Framework for Building Contextual