<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="joshi-etal-2026-ecostatkg">
    <titleInfo>
        <title>EcoStatKG: A Domain-Specific Statistical Knowledge Graph and Benchmark for Evaluating Numerical Hallucination in LLMs</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Ratnesh</namePart>
        <namePart type="given">Kumar</namePart>
        <namePart type="family">Joshi</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Sagnik</namePart>
        <namePart type="family">Sengupta</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Asif</namePart>
        <namePart type="family">Ekbal</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 1: Long 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">979-8-89176-439-2</identifier>
    </relatedItem>
    <abstract>Large language models routinely produce numerical hallucinations—plausible but incorrect quantities, percentages, and statistics—yet existing benchmarks focus almost exclusively on entity-level factuality. We introduce EcoStatKG, a knowledge graph (KG) of 50,746 statistical triples and 306,231 structural triples from 6,871 Wikipedia articles on environmental sustainability, and NumHallu, a 500-question numerical benchmark grounded in held-out KG topics. We use these resources to diagnose how knowledge-base coverage affects eight retrieval strategies across six LLMs. In the full-index, oracle-coverage condition, simple cosine retrieval reaches 0.989 Numerical Exact Match (NEM), statistically indistinguishable from Graph RAG (0.992; paired-bootstrap p=0.14). In the deliberately constructed held-out condition, which removes answer-bearing triples, retrieval methods fall below the zero-shot baseline (0.506 NEM), exposing the risk that topically related numerical distractors can override parametric knowledge. Similarity-threshold filtering restores KG Cosine to near-baseline performance (0.511), rather than establishing a new state-of-the-art. A bounded clinical-domain proof of concept reproduces the coverage and distractor patterns. The contribution is a diagnostic study and resource: its results clarify when numerical retrieval helps, when missing coverage can make it harmful, and why confidence gating is best viewed as damage control.</abstract>
    <identifier type="citekey">joshi-etal-2026-ecostatkg</identifier>
    <location>
        <url>https://aclanthology.org/2026.aacl-long.119/</url>
    </location>
    <part>
        <date>2026-11</date>
        <extent unit="page">
            <start>2598</start>
            <end>2621</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T EcoStatKG: A Domain-Specific Statistical Knowledge Graph and Benchmark for Evaluating Numerical Hallucination in LLMs
%A Joshi, Ratnesh Kumar
%A Sengupta, Sagnik
%A Ekbal, Asif
%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 1: Long Papers)
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-439-2
%F joshi-etal-2026-ecostatkg
%X Large language models routinely produce numerical hallucinations—plausible but incorrect quantities, percentages, and statistics—yet existing benchmarks focus almost exclusively on entity-level factuality. We introduce EcoStatKG, a knowledge graph (KG) of 50,746 statistical triples and 306,231 structural triples from 6,871 Wikipedia articles on environmental sustainability, and NumHallu, a 500-question numerical benchmark grounded in held-out KG topics. We use these resources to diagnose how knowledge-base coverage affects eight retrieval strategies across six LLMs. In the full-index, oracle-coverage condition, simple cosine retrieval reaches 0.989 Numerical Exact Match (NEM), statistically indistinguishable from Graph RAG (0.992; paired-bootstrap p=0.14). In the deliberately constructed held-out condition, which removes answer-bearing triples, retrieval methods fall below the zero-shot baseline (0.506 NEM), exposing the risk that topically related numerical distractors can override parametric knowledge. Similarity-threshold filtering restores KG Cosine to near-baseline performance (0.511), rather than establishing a new state-of-the-art. A bounded clinical-domain proof of concept reproduces the coverage and distractor patterns. The contribution is a diagnostic study and resource: its results clarify when numerical retrieval helps, when missing coverage can make it harmful, and why confidence gating is best viewed as damage control.
%U https://aclanthology.org/2026.aacl-long.119/
%P 2598-2621

Markdown (Informal)

[EcoStatKG: A Domain-Specific Statistical Knowledge Graph and Benchmark for Evaluating Numerical Hallucination in LLMs](https://aclanthology.org/2026.aacl-long.119/) (Joshi et al., AACL-IJCNLP 2026)

• EcoStatKG: A Domain-Specific Statistical Knowledge Graph and Benchmark for Evaluating Numerical Hallucination in LLMs (Joshi et al., AACL-IJCNLP 2026)

ACL

• Ratnesh Kumar Joshi, Sagnik Sengupta, and Asif Ekbal. 2026. EcoStatKG: A Domain-Specific Statistical Knowledge Graph and Benchmark for Evaluating Numerical Hallucination in LLMs. In 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 1: Long Papers), pages 2598–2621, Hengqin, China. Association for Computational Linguistics.