@inproceedings{al-khalifa-2026-twenty,
title = "Twenty-Five Years of Inter-Annotator Agreement in {NLP}: A Systematic Survey of Trends, Metrics, and Reporting Practices",
author = "Al-Khalifa, Hend",
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 1: Long Papers)",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.aacl-long.27/",
pages = "579--602",
ISBN = "979-8-89176-439-2",
abstract = "Large language models are now deployed as annotators across NLP, generating training labels and evaluation scores, yet the field has no established standard for verifying whether LLM-generated labels are reliable, internally consistent, or comparable across studies. This methodological gap does not begin with LLMs: a systematic corpus study of 665 ACL Anthology papers from 1999 to 2026 reveals that even classical human annotation in NLP suffers from systematic measurement and reporting failures that have left a growing body of work scientifically incomparable. Papers were identified through keyword search and organised via topic modelling into a thematic taxonomy spanning eight annotation domains from named entity recognition to disagreement modelling. Full-text analysis of 549 papers yields three principal findings. First, Cohen{'}s kappa dominates across all task types (259 papers, 38.9{\%}), regardless of whether its assumptions are satisfied; ICC, appropriate for continuous-scale tasks such as MT evaluation, appears in only 16 full-text papers (2.4{\%}). Second, 226 of the 665 papers (34.0{\%}) name no inter-annotator agreement (IAA) coefficient anywhere in the full text, which is a methodology-reporting concern; a further 50.1{\%} of papers name a coefficient in the body but not in the title or abstract, so 84.1{\%} of papers omit the metric from the abstract layer. We frame the body-only gap as a discoverability concern (the methodology is reported but is not visible at the abstract layer readers and reviewers encounter first), distinct from the full-text omission. Third, annotator disagreement modelling and LLM-as-annotator constitute the fastest-growing area of the corpus (124 papers, 18.6{\%}): 26 of these appeared between 1999 and 2022 and 98 between 2023 and 2026, the only dimension whose volume rises in that window. Fewer than 5{\%} of papers report confidence intervals. We provide a large-scale empirical analysis of metric reporting practice within the ACL Anthology and a seven-point agenda for standardised reporting, complementing prior normative guidance with field-scale descriptive evidence."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="al-khalifa-2026-twenty">
<titleInfo>
<title>Twenty-Five Years of Inter-Annotator Agreement in NLP: A Systematic Survey of Trends, Metrics, and Reporting Practices</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hend</namePart>
<namePart type="family">Al-Khalifa</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 are now deployed as annotators across NLP, generating training labels and evaluation scores, yet the field has no established standard for verifying whether LLM-generated labels are reliable, internally consistent, or comparable across studies. This methodological gap does not begin with LLMs: a systematic corpus study of 665 ACL Anthology papers from 1999 to 2026 reveals that even classical human annotation in NLP suffers from systematic measurement and reporting failures that have left a growing body of work scientifically incomparable. Papers were identified through keyword search and organised via topic modelling into a thematic taxonomy spanning eight annotation domains from named entity recognition to disagreement modelling. Full-text analysis of 549 papers yields three principal findings. First, Cohen’s kappa dominates across all task types (259 papers, 38.9%), regardless of whether its assumptions are satisfied; ICC, appropriate for continuous-scale tasks such as MT evaluation, appears in only 16 full-text papers (2.4%). Second, 226 of the 665 papers (34.0%) name no inter-annotator agreement (IAA) coefficient anywhere in the full text, which is a methodology-reporting concern; a further 50.1% of papers name a coefficient in the body but not in the title or abstract, so 84.1% of papers omit the metric from the abstract layer. We frame the body-only gap as a discoverability concern (the methodology is reported but is not visible at the abstract layer readers and reviewers encounter first), distinct from the full-text omission. Third, annotator disagreement modelling and LLM-as-annotator constitute the fastest-growing area of the corpus (124 papers, 18.6%): 26 of these appeared between 1999 and 2022 and 98 between 2023 and 2026, the only dimension whose volume rises in that window. Fewer than 5% of papers report confidence intervals. We provide a large-scale empirical analysis of metric reporting practice within the ACL Anthology and a seven-point agenda for standardised reporting, complementing prior normative guidance with field-scale descriptive evidence.</abstract>
<identifier type="citekey">al-khalifa-2026-twenty</identifier>
<location>
<url>https://aclanthology.org/2026.aacl-long.27/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>579</start>
<end>602</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Twenty-Five Years of Inter-Annotator Agreement in NLP: A Systematic Survey of Trends, Metrics, and Reporting Practices
%A Al-Khalifa, Hend
%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 al-khalifa-2026-twenty
%X Large language models are now deployed as annotators across NLP, generating training labels and evaluation scores, yet the field has no established standard for verifying whether LLM-generated labels are reliable, internally consistent, or comparable across studies. This methodological gap does not begin with LLMs: a systematic corpus study of 665 ACL Anthology papers from 1999 to 2026 reveals that even classical human annotation in NLP suffers from systematic measurement and reporting failures that have left a growing body of work scientifically incomparable. Papers were identified through keyword search and organised via topic modelling into a thematic taxonomy spanning eight annotation domains from named entity recognition to disagreement modelling. Full-text analysis of 549 papers yields three principal findings. First, Cohen’s kappa dominates across all task types (259 papers, 38.9%), regardless of whether its assumptions are satisfied; ICC, appropriate for continuous-scale tasks such as MT evaluation, appears in only 16 full-text papers (2.4%). Second, 226 of the 665 papers (34.0%) name no inter-annotator agreement (IAA) coefficient anywhere in the full text, which is a methodology-reporting concern; a further 50.1% of papers name a coefficient in the body but not in the title or abstract, so 84.1% of papers omit the metric from the abstract layer. We frame the body-only gap as a discoverability concern (the methodology is reported but is not visible at the abstract layer readers and reviewers encounter first), distinct from the full-text omission. Third, annotator disagreement modelling and LLM-as-annotator constitute the fastest-growing area of the corpus (124 papers, 18.6%): 26 of these appeared between 1999 and 2022 and 98 between 2023 and 2026, the only dimension whose volume rises in that window. Fewer than 5% of papers report confidence intervals. We provide a large-scale empirical analysis of metric reporting practice within the ACL Anthology and a seven-point agenda for standardised reporting, complementing prior normative guidance with field-scale descriptive evidence.
%U https://aclanthology.org/2026.aacl-long.27/
%P 579-602
Markdown (Informal)
[Twenty-Five Years of Inter-Annotator Agreement in NLP: A Systematic Survey of Trends, Metrics, and Reporting Practices](https://aclanthology.org/2026.aacl-long.27/) (Al-Khalifa, AACL-IJCNLP 2026)
ACL