@inproceedings{pinto-figueira-2026-auditing,
title = "Auditing Bias Detection Under Source Shift: A Source-Aware Benchmark and Evaluation Protocol",
author = "Pinto, Ana and
Figueira, Alvaro",
editor = "Zhang, Yue and
Xu, Ruifeng and
Purwarianti, Ayu and
Rong, Tong and
Suzuki, Jun",
booktitle = "Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the {A}sia-Pacific Chapter of the {A}ssociation for {C}omputational {L}inguistics (Volume 2: Short Papers)",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.aacl-short.55/",
pages = "854--878",
ISBN = "979-8-89176-440-8",
abstract = "Bias-detection benchmarks routinely report accuracies above 90{\%}, but a high number can reflect where a sentence came from as much as whether it contains a stereotype or an unfair claim about a group. We treat this as an evaluation problem and propose an integrated auditing protocol that makes source confounding measurable: an LLM-assisted label audit with cross-family verification, pair-aware splitting with role tracking, source-aware evaluation with a source-only baseline on the actual test split, counterfactual discrimination, and an artifact analysis of edited text. We apply it to a three-source corpus of 10,304 sentences that we audit and release. The audit reduces the global source-only majority baseline from 95.0{\%} to 55.4{\%}, and the same protocol shows that the real-only test split still carries a 77.6{\%} source-only baseline, so confounding is reduced and measured but far from eliminated. Against that baseline, fine-tuned BERT and GPT-2 gain roughly 18 pp and outperform attention-derived models; four augmentation variants that look ordered in the mean fall within seed variance under paired testing; and a real-vs-edited probe separates Gemini counterfactuals from real text, although the signal disappears once the gold label is held fixed and has no measurable accuracy effect. Applied to CrowS-Pairs, the label audit flags 33.2{\%} of the supposedly neutral members as still containing a group-level stereotype. The main contribution is the protocol, which exposes confounding that standard pipelines would report as results."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="pinto-figueira-2026-auditing">
<titleInfo>
<title>Auditing Bias Detection Under Source Shift: A Source-Aware Benchmark and Evaluation Protocol</title>
</titleInfo>
<name type="personal">
<namePart type="given">Ana</namePart>
<namePart type="family">Pinto</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alvaro</namePart>
<namePart type="family">Figueira</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 2: Short Papers)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruifeng</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ayu</namePart>
<namePart type="family">Purwarianti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tong</namePart>
<namePart type="family">Rong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jun</namePart>
<namePart type="family">Suzuki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Hengqin, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-440-8</identifier>
</relatedItem>
<abstract>Bias-detection benchmarks routinely report accuracies above 90%, but a high number can reflect where a sentence came from as much as whether it contains a stereotype or an unfair claim about a group. We treat this as an evaluation problem and propose an integrated auditing protocol that makes source confounding measurable: an LLM-assisted label audit with cross-family verification, pair-aware splitting with role tracking, source-aware evaluation with a source-only baseline on the actual test split, counterfactual discrimination, and an artifact analysis of edited text. We apply it to a three-source corpus of 10,304 sentences that we audit and release. The audit reduces the global source-only majority baseline from 95.0% to 55.4%, and the same protocol shows that the real-only test split still carries a 77.6% source-only baseline, so confounding is reduced and measured but far from eliminated. Against that baseline, fine-tuned BERT and GPT-2 gain roughly 18 pp and outperform attention-derived models; four augmentation variants that look ordered in the mean fall within seed variance under paired testing; and a real-vs-edited probe separates Gemini counterfactuals from real text, although the signal disappears once the gold label is held fixed and has no measurable accuracy effect. Applied to CrowS-Pairs, the label audit flags 33.2% of the supposedly neutral members as still containing a group-level stereotype. The main contribution is the protocol, which exposes confounding that standard pipelines would report as results.</abstract>
<identifier type="citekey">pinto-figueira-2026-auditing</identifier>
<location>
<url>https://aclanthology.org/2026.aacl-short.55/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>854</start>
<end>878</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Auditing Bias Detection Under Source Shift: A Source-Aware Benchmark and Evaluation Protocol
%A Pinto, Ana
%A Figueira, Alvaro
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 2: Short Papers)
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-440-8
%F pinto-figueira-2026-auditing
%X Bias-detection benchmarks routinely report accuracies above 90%, but a high number can reflect where a sentence came from as much as whether it contains a stereotype or an unfair claim about a group. We treat this as an evaluation problem and propose an integrated auditing protocol that makes source confounding measurable: an LLM-assisted label audit with cross-family verification, pair-aware splitting with role tracking, source-aware evaluation with a source-only baseline on the actual test split, counterfactual discrimination, and an artifact analysis of edited text. We apply it to a three-source corpus of 10,304 sentences that we audit and release. The audit reduces the global source-only majority baseline from 95.0% to 55.4%, and the same protocol shows that the real-only test split still carries a 77.6% source-only baseline, so confounding is reduced and measured but far from eliminated. Against that baseline, fine-tuned BERT and GPT-2 gain roughly 18 pp and outperform attention-derived models; four augmentation variants that look ordered in the mean fall within seed variance under paired testing; and a real-vs-edited probe separates Gemini counterfactuals from real text, although the signal disappears once the gold label is held fixed and has no measurable accuracy effect. Applied to CrowS-Pairs, the label audit flags 33.2% of the supposedly neutral members as still containing a group-level stereotype. The main contribution is the protocol, which exposes confounding that standard pipelines would report as results.
%U https://aclanthology.org/2026.aacl-short.55/
%P 854-878
Markdown (Informal)
[Auditing Bias Detection Under Source Shift: A Source-Aware Benchmark and Evaluation Protocol](https://aclanthology.org/2026.aacl-short.55/) (Pinto & Figueira, AACL-IJCNLP 2026)
ACL