The DOI matches its Crossref record; 106 of the 106 references it cites were found in a registry or an index; none of the 74 sources its claims cite is checked for support yet.
Best practices for the human evaluation of automatically generated text
Open accessAbstract
Currently, there is little agreement as to how Natural Language Generation (NLG) systems should be evaluated, with a particularly high degree of variation in the way that human evaluation is carried out.This paper provides an overview of how human evaluation is currently conducted, and presents a set of best practices, grounded in the literature.With this paper, we hope to contribute to the quality and consistency of human evaluations in NLG.
Overview
CoChat Summary
TL;DR
This paper surveys current human evaluation practices in NLG and proposes best practices grounded in NLG, statistics, and behavioral sciences. It finds high variation and underreporting in evaluation design, sampling, agreement, and statistics, and recommends more standardized, reader-focused, well-reported evaluations.
Key findings
- In a survey of INLG (N=51) and ACL (N=38) 2018 papers, 80% of empirical papers reported automatic metrics, up from 38.2% in 2005–2014, while only 3% presented an extrinsic evaluation, down from nearly 25% in the earlier survey.
- Human evaluation reporting is inconsistent: median 4 annotators; 55% specified participant numbers; 18% reported demographics; 12.5% reported inter-annotator agreement; median 100 items (range 2–5,400); 83% of papers reporting these figures had all annotators see all examples.
- Rating practices vary: 5-point Likert is most popular, preference ratings second; only 34% reported the number of ratings per criterion; 33% reported statistical analyses; 19% explicitly stated hypotheses or research questions.
- Recommended best practices include using separate, well-defined criteria; preferring large-scale reader-focused designs; reporting sample size and demographics; using multiple-item 7-point Likert scales or continuous ranking (with TrueSkill for multiple systems); counterbalancing/randomizing order; and distinguishing exploratory from confirmatory analysis, with preregistration for confirmatory studies.
Methods
- Bibliometric overview of human evaluation practices in papers published at INLG (N=51) and ACL (N=38) in 2018, merged because no noticeable differences were observed; ACL tracks examined were Machine Translation, Summarization, Question Answering, and Generation.
- Narrative synthesis of prior NLG, statistics, and behavioral-science literature to derive best practices, mostly restricted to intrinsic evaluation.
- No new human-subjects experiment; recommendations are summarized in Table 3.
Limitations
- Stated: the paper covers a broad topic and, for reasons of space, could not cover all aspects of human evaluation in detail; future work aims to provide a more extensive set of best practices.
- Stated/evident: the survey is limited to INLG and ACL 2018 papers, and many papers did not report participant numbers, demographics, inter-annotator agreement, design details, or statistics, so the observed numbers may not reflect reality.
- Evident: the best practices mostly restrict themselves to intrinsic evaluation, and some recommendations—such as whether multiple-item Likert scales make reliability comparable to ranking tasks—have not been empirically tested in NLG.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
0 Valid · 5 Not valid · 55 Not checked · 47 Own results and interpretations
A result the paper states — it cites no source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“A minority (33%) of papers report one or more statistical analyses for their human evaluation to investigate if findings are statistically significant.”
§ Best practices for the human evaluation of automatically generated text > 3.6 Statistics and data analysis
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Only 34% of papers with a human evaluation reported the number of ratings to measure a criterion.”
§ Best practices for the human evaluation of automatically generated text > 3.5 Number of questions and types of scales
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper states — it cites no source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“And conversely, some papers reporting hypotheses do not perform a statistical test.”
§ Best practices for the human evaluation of automatically generated text > 3.6 Statistics and data analysis
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Among papers that reported these numbers, we observed a median of 100 items used for human evaluation in INLG and ACL papers.”
§ Best practices for the human evaluation of automatically generated text > 3.4 Design
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper states — it cites no source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“One obstacle to addressing the difficulties identified in this section is the lack of a standardised nomenclature for different text quality criteria.”
§ Best practices for the human evaluation of automatically generated text > 4.1 Text quality and criteria
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
50 more claims
Create a free account to see all 107 claims and to verify the 74 references still open.
- It may state two separate claims as one. 27
- Probably not a substantial claim. 30
- The quote only partly states this claim. 1
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Such models incorporate variation in the quality of generated instances, annotator responses, and annotator sensitivity, and are recommended by van der Lee et al. (2019) for analyzing human evaluations.”
From With Little Power Comes Great Responsibility (2020) · cites this paper as van der Lee et al. (2019)
Experimental — accuracy not yet measuredThe cited source says what the claim attributes to it.
In the citing paper
“Recognizing this, best practices often involve multiple human annotators collaborating in the evaluation (Van Der Lee et al., 2019).”
From ChatEval: Towards Better LLM-based Evaluators through Multi-Agent Debate (2023) · cites this paper as Van Der Lee et al., 2019
The cited source says what the claim attributes to it.
In the citing paper
“The evaluation of the quality of machine-generated excerpts has traditionally relied on human judgement, which is considered the gold standard for open-domain generation systems [van der Lee et al., 2019;Gehrmann et al., 2019a].”
From Authorship Attribution in the Era of LLMs: Problems, Methodologies, and Challenges (2025) · cites this paper as van der Lee et al., 2019
The cited source says what the claim attributes to it.
In the citing paper
“Past work has reported great variation in how human evaluations are done (van der Lee et al., 2019).”
From With Little Power Comes Great Responsibility (2020) · cites this paper as van der Lee et al., 2019
The cited source says what the claim attributes to it.
In the citing paper
“But, as similarly found by van der Lee et al. (2019), many key details were not reported in this sample of experiments.”
From With Little Power Comes Great Responsibility (2020) · cites this paper as van der Lee et al. (2019)
1 more claim about this paper
Create a free account to see every claim other papers make about this one.
Citations
Cites
Statistical power analysis for the behavioral sciences
(1990) · Computers, Environment and Urban Systems · doi:10.1016/0198-9715(90)90050-4 · Cited by 65,269
Statistical Power Analysis for the Behavioral Sciences
Cohen, J. (2013) · doi:10.4324/9780203771587 · Cited by 25,373
Interrater reliability: The kappa statistic
McHugh, M. L. (2012) · Biochemia Medica · doi:10.11613/bm.2012.031 · Cited by 19,877 · cited 3×
“An additional consideration is that some statistics (such as κ ; see McHugh, 2012) make overly restrictive assumptions, though they have the advantage of accounting for chance agreement.”
BackgroundInfluentialThe magical number seven, plus or minus two: Some limits on our capacity for processing information.
Miller, G. A. (1956) · Psychological Review · doi:10.1037/h0043158 · Cited by 17,812 · cited 1×
“Most of the experimental literature’s findings found that 7-point scales maximise reliability, validity and discriminative power (for instance, Miller, 1956; Green and Rao, 1970; Jones, 1968; Cicchetti et al., 1985; Lissitz and Green, 1975; Preston and Colman, 2000).”
BackgroundThe magical number seven, plus or minus two: Some limits on our capacity for processing information.
Miller, G. A. (1994) · Psychological Review · doi:10.1037/0033-295x.101.2.343 · Cited by 17,161
The weirdest people in the world?
Henrich, J., Heine, S. J., & Norenzayan, A. (2010) · Behavioral and Brain Sciences · doi:10.1017/s0140525x0999152x · Cited by 12,641
Journal of personality and social psychology
Armstrong, J. S. (1988) · International Journal of Forecasting · doi:10.1016/0169-2070(88)90126-4 · Cited by 10,485
Random effects structure for confirmatory hypothesis testing: Keep it maximal
Barr, D. J., Levy, R., Scheepers, C., & Tily, H. J. (2013) · Journal of Memory and Language · doi:10.1016/j.jml.2012.11.001 · Cited by 10,328 · cited 1×
“Mixed effects models capture such potential item aspects very well (e.g., Barr et al., 2013), and deserve more attention in NLG. Finally, Bayesian models are worth exploring, because they are less sensitive to the aforementioned problems with NHST (e.g., Gelman et al., 2006; Wagenmakers, 2007).”
BackgroundPower failure: why small sample size undermines the reliability of neuroscience
Button, K. S., Ioannidis, J. P. A., Mokrysz, C., Nosek, B. A., Flint, J., Robinson, E. S. J., & Munafò, M. R. (2013) · Nature Reviews Neuroscience · doi:10.1038/nrn3475 · Cited by 7,641
Simmons, J. P., Nelson, L. D., & Simonsohn, U. (2011) · Psychological Science · doi:10.1177/0956797611417632 · Cited by 6,567 · cited 2×
“Conducting and analysing a human experiment is like entering a garden of forking paths (Gelman and Loken, 2013): along the way researchers have many choices to make, and even though each choice may be small and seemingly innocuous, collectively they can have a substantial effect on the outcome of the statistical analy…”
Background
Research works printed but not linked: 6
- Thiago Castro Ferreira, Chris van der Lee, Emiel van Miltenburg, and Emiel Krahmer. 2019. Neural data- to-text generation: A comparison between pipeline and end-to-end architectures. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing, Hong Kong, SAR. Associa- tion for Computational Linguistics.
- Jekaterina Novikova, Ondřej Dušek, and Verena Rieser. 2018. RankME: Reliable human ratings for Natural Language Generation. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 72-78. · DOI
- Jan Renkema. 2012. Schrijfwijzer, 5 edition. SDU Uit- gevers, Den Haag, The Netherlands.
- Donia Scott and Johanna Moore. 2007. An NLG eval- uation competition? Eight reasons to be cautious. In Proceedings of the Workshop on Shared Tasks and Comparative Evaluation in Natural Language Gen- eration, pages 22-23.
- Renske Van Enschot, Wilbert Spooren, Antal van den Bosch, Christian Burgers, Liesbeth Degand, Jacque- line Evers-Vermeul, Florian Kunneman, Christine Liebrecht, Yvette Linders, and Alfons Maes. 2017. Taming our wild data: On intercoder reliability in discourse research. Unpublished Manuscript, pages 1-18.
- David L Vannette. 2016. Testing the effects of differ- ent types of attention interventions on data quality in web surveys. Experimental evidence from a 14 country study. In 71st Annual Conference of the American Association for Public Opinion Research.
Cited by
A Survey on Evaluation of Large Language Models
Chang, Y., Wang, X., Wang, J., Wu, Y., Yang, L., Zhu, K., Chen, H., Yi, X., Wang, C., Wang, Y., et al. (2024) · ACM Transactions on Intelligent Systems and Technology · doi:10.1145/3641289 · Cited by 4,254
Lee, M., Liang, P., & Yang, Q. (2022) · CHI Conference on Human Factors in Computing Systems · doi:10.1145/3491102.3502030 · Cited by 345
“Then, we describe potential use cases of CoAuthor in HCI and NLP research, exemplifying the potential of CoAuthor serving as a boundary object.”
BackgroundInfluentialAll That’s ‘Human’ Is Not Gold: Evaluating Human Evaluation of Generated Text
Clark, E., August, T., Serrano, S., Haduong, N., Gururangan, S., & Smith, N. A. (2021) · Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers) · doi:10.18653/v1/2021.acl-long.565 · Cited by 290
“Though evaluations with end-users in an applied setting are encouraged (Belz and Reiter, 2006), in practice, most human evaluations instead ask people to rate generated text’s intrinsic quality (van der Lee et al., 2019; Howcroft et al., 2020).”
BackgroundMethodAll That's 'Human' Is Not Gold: Evaluating Human Evaluation of Generated Text
Clark, E., August, T., Serrano, S., Haduong, N., Gururangan, S., & Smith, N. A. (2021) · arXiv · doi:10.48550/arxiv.2107.00061 · Cited by 290
Automatic Detection of Generated Text is Easiest when Humans are Fooled
Ippolito, D., Duckworth, D., Callison-Burch, C., & Eck, D. (2020) · Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics · doi:10.18653/v1/2020.acl-main.164 · Cited by 264
“The related (but not identical) task of asking human raters to judge the quality of machine-generated excerpts remains the gold-standard for evaluating open-domain generation systems (van der Lee et al., 2019).”
Background
Metadata
- Authors
- van der Lee, C., Gatt, A., van Miltenburg, E., Wubben, S., & Krahmer, E.
- Affiliations
- Tilburg University (NL)
- ORCID
- Krahmer, Emiel
- Venue
- Proceedings of the 12th International Conference on Natural Language Generation
- Published
- 2019
- Pages
- 355-368
- OpenAlex
- W2992347006
- Semantic Scholar
- d1aa325a5adefeba786d4c50a8ca9a8df9598f32
- Cited by
- 205
- References
- 103
- Citations by year
- 2026: 23 · 2025: 23 · 2024: 27 · 2023: 23 · 2022: 34 · 2021: 37 · 2020: 38
- FWCI
- 10.43
- Type
- conference paper
- Language
- en
- Open access
- hybrid
- doi.org
- Licence
- cc-by
- Locations
- International Conference on Natural Language Generation (publishedVersion, cc-by)
- Research portal (Tilburg University) (submittedVersion, other-oa)
- Tilburg University Research Portal (submittedVersion, other-oa)
- Data Archiving and Networked Services (DANS) (submittedVersion, other-oa)
- Funders
- Nederlandse Organisatie voor Wetenschappelijk Onderzoek (360-89-050)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
- Physical Sciences › Computer Science › Artificial Intelligence › Speech and dialogue systems
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Keywords
- natural language generation, human evaluation, evaluation best practices, text generation, assessment consistency, assessment quality
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
A language model drew the claims from the paper on 2026-10-05, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-05 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Best practices for the human evaluation of automatically generated text. https://app.cochat.ai/papers/10.18653/v1/w19-8643
- The paper
- van der Lee, C., Gatt, A., van Miltenburg, E., Wubben, S., & Krahmer, E. (2019). Best practices for the human evaluation of automatically generated text. Proceedings of the 12th International Conference on Natural Language Generation. https://doi.org/10.18653/v1/w19-8643
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.