arXiv cs.CL· Yuchen Miao, Zijun Wang, Chang Han, Yurui Shi, Mingtai Zhang, Siyang Xu·· 9 天前AI 评分22
MARS-Gov：面向荷兰政府文档官僚偏见检测的开集立场筛查框架
The "10th Juror": Open-Set Standpoint Screening for Bureaucratic Bias Detection
AI 导读
MARS-Gov 是一个立场感知的多智能体框架，结合法律检索、开集目标筛查、专业陪审员、保守路由与改写验证，用于荷兰政府文档的闭环偏见治理。在 DGDB 上以 0.880 F1 刷新 SOTA，超过最强零样本 LLM 检测器 20.2 分（相对提升 29.8%）、最佳监督式荷兰语编码器 6.8 分，并将不必要干预降至 2.5%。
正文 · 原文
Abstract:Presupposing the boundaries of bias is itself a form of bias. We study closed-loop bias governance for Dutch government documents, where a system must detect biased language, ground decisions in legal and contextual evidence, rewrite problematic sentences when intervention is warranted, and verify that the rewrite mitigates harm without distorting meaning. Existing methods face three challenges: (i) discriminative classifiers capture surface regularities but lack normative grounding; (ii) zero-shot LLMs often adopt generic viewpoints and over-flag ambiguous administrative language; and (iii) fixed taxonomies inherit the Closed-World Assumption, missing emerging local targets. We propose MARS-Gov, a standpoint-aware multi-agent framework that combines legal retrieval, open-set target screening, specialized jurors, conservative routing, and rewrite verification. When screening finds an uncovered group, MARS-Gov instantiates a dynamic "10th juror" to deliberate outside the fixed panel. On DGDB, MARS-Gov sets a new SOTA with 0.880 F1, outperforming the strongest zero-shot LLM detector by 20.2 points (29.8% relative) and the best supervised Dutch encoder by 6.8 points, while reducing unnecessary interventions to 2.5%. Leave-One-Category-Out (LOCO) evaluation recovers held-out categories with 85.1% Correct@1 and 93.8% Correct@3.
| Comments: | 19 pages, 6 figures. Accepted at EMNLP 2026 |
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2610.11136 [cs.CL] |
| (or arXiv:2610.11136v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.11136 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Zijun Wang [view email]
[v1]
Thu, 8 Oct 2026 03:05:08 UTC (924 KB)
来源：arXiv cs.CL · arxiv.org