<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="ge-etal-2026-trustrag">
    <titleInfo>
        <title>TrustRAG: Metadata-Governed RAG for Compliance-Oriented Question Answering</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Huiling</namePart>
        <namePart type="family">Ge</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Zhiguo</namePart>
        <namePart type="family">Shi</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Yungui</namePart>
        <namePart type="family">Nie</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <originInfo>
        <dateIssued>2026-10</dateIssued>
    </originInfo>
    <typeOfResource>text</typeOfResource>
    <relatedItem type="host">
        <titleInfo>
            <title>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</title>
        </titleInfo>
        <name type="personal">
            <namePart type="given">Hung-Yi</namePart>
            <namePart type="family">Lee</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Sunipa</namePart>
            <namePart type="family">Dev</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Jessy</namePart>
            <namePart type="family">Li</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Desmond</namePart>
            <namePart type="family">Elliott</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <originInfo>
            <publisher>Association for Computational Linguistics</publisher>
            <place>
                <placeTerm type="text">Budapest, Hungary</placeTerm>
            </place>
        </originInfo>
        <genre authority="marcgt">conference publication</genre>
        <identifier type="isbn">979-8-89176-459-0</identifier>
    </relatedItem>
    <abstract>Retrieval-Augmented Generation (RAG) enhances large language models with external evidence, but in compliance-oriented QA, semantic relevance alone does not determine whether evidence should be used for generation. A retrieved document may be relevant and factually correct, yet inadmissible because it is inaccessible to the current user, explicitly superseded, or in unresolved conflict with other active evidence. We formalize three post-retrieval governance failures in this setting: unauthorized leakage, stale citation, and unsupported answers. To address them, we propose TrustRAG, a RAG framework that incorporates a metadata-governed routing layer between the retriever and the generator. Given correct and complete governance metadata for the retrieved candidate set, this layer filters inadmissible evidence before generation, returns a deterministic refusal when no admissible evidence remains, and preserves unresolved conflicts for conflict-reporting generation. We construct AeroTrustBench, a counterfactual compliance QA benchmark for evaluating evidence governance in RAG. Across Qwen-Max, DeepSeek-V3, and GLM-4-32B-0414, TrustRAG consistently reduces the three failures relative to unguided RAG and baselines based on prompting, filtering, routing, verification, and recency. Mechanism-matched and 50-document comparisons show benefits beyond filtering under larger candidate pools, while metadata-noise tests and human audits characterize dependencies on upstream metadata and evaluator choice.</abstract>
    <identifier type="citekey">ge-etal-2026-trustrag</identifier>
    <location>
        <url>https://aclanthology.org/2026.emnlp-main.1533/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>35002</start>
            <end>35020</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T TrustRAG: Metadata-Governed RAG for Compliance-Oriented Question Answering
%A Ge, Huiling
%A Shi, Zhiguo
%A Nie, Yungui
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F ge-etal-2026-trustrag
%X Retrieval-Augmented Generation (RAG) enhances large language models with external evidence, but in compliance-oriented QA, semantic relevance alone does not determine whether evidence should be used for generation. A retrieved document may be relevant and factually correct, yet inadmissible because it is inaccessible to the current user, explicitly superseded, or in unresolved conflict with other active evidence. We formalize three post-retrieval governance failures in this setting: unauthorized leakage, stale citation, and unsupported answers. To address them, we propose TrustRAG, a RAG framework that incorporates a metadata-governed routing layer between the retriever and the generator. Given correct and complete governance metadata for the retrieved candidate set, this layer filters inadmissible evidence before generation, returns a deterministic refusal when no admissible evidence remains, and preserves unresolved conflicts for conflict-reporting generation. We construct AeroTrustBench, a counterfactual compliance QA benchmark for evaluating evidence governance in RAG. Across Qwen-Max, DeepSeek-V3, and GLM-4-32B-0414, TrustRAG consistently reduces the three failures relative to unguided RAG and baselines based on prompting, filtering, routing, verification, and recency. Mechanism-matched and 50-document comparisons show benefits beyond filtering under larger candidate pools, while metadata-noise tests and human audits characterize dependencies on upstream metadata and evaluator choice.
%U https://aclanthology.org/2026.emnlp-main.1533/
%P 35002-35020

Markdown (Informal)

[TrustRAG: Metadata-Governed RAG for Compliance-Oriented Question Answering](https://aclanthology.org/2026.emnlp-main.1533/) (Ge et al., EMNLP 2026)

• TrustRAG: Metadata-Governed RAG for Compliance-Oriented Question Answering (Ge et al., EMNLP 2026)

ACL

• Huiling Ge, Zhiguo Shi, and Yungui Nie. 2026. TrustRAG: Metadata-Governed RAG for Compliance-Oriented Question Answering. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 35002–35020, Budapest, Hungary. Association for Computational Linguistics.