Missing Information, Unresponsive Authors, Experimental Flaws: The Impossibility of Assessing the Reproducibility of Previous Human Evaluations in NLP

Anya Belz; Craig Thomson; Ehud Reiter; Gavin Abercrombie; Jose M. Alonso-Moral; Mohammad Arvan; Jackie Cheung; Mark Cieliebak; Elizabeth Clark; Kees van Deemter; Tanvi Dinkar; Ondřej Dušek; Steffen Eger; Qixiang Fang; Albert Gatt; Dimitra Gkatzia; Javier González-Corbelle; Dirk Hovy; Manuela Hürlimann; Takumi Ito; John D. Kelleher; Filip Klubicka; Huiyuan Lai; Chris van der Lee; Emiel van Miltenburg; Yiru Li; Saad Mahamood; Margot Mieskes; Malvina Nissim; Natalie Parde; Ondřej Plátek; Verena Rieser; Pablo Mosteiro Romero; Joel Tetreault; Antonio Toral; Xiaojun Wan; Leo Wanner; Lewis Watson; Diyi Yang

Missing Information, Unresponsive Authors, Experimental Flaws: The Impossibility of Assessing the Reproducibility of Previous Human Evaluations in NLP

Anya Belz, Craig Thomson, Ehud Reiter, Gavin Abercrombie, Jose M. Alonso-Moral, Mohammad Arvan, Jackie Cheung, Mark Cieliebak, Elizabeth Clark, Kees van Deemter, Tanvi Dinkar, Ondřej Dušek, Steffen Eger, Qixiang Fang, Albert Gatt, Dimitra Gkatzia, Javier González-Corbelle, Dirk Hovy, Manuela Hürlimann, Takumi ItoJohn D. Kelleher, Filip Klubicka, Huiyuan Lai, Chris van der Lee, Emiel van Miltenburg, Yiru Li, Saad Mahamood, Margot Mieskes, Malvina Nissim, Natalie Parde, Ondřej Plátek, Verena Rieser, Pablo Mosteiro Romero, Joel Tetreault, Antonio Toral, Xiaojun Wan, Leo Wanner, Lewis Watson, Diyi Yang

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › Academic › peer-review

Abstract

We report our efforts in identifying a set of previous human evaluations in NLP that would be suitable for a coordinated study examining what makes human evaluations in NLP more/less reproducible. We present our results and findings, which include that just 13% of papers had (i) sufficiently low barriers to reproduction, and (ii) enough obtainable information, to be considered for reproduction, and that all but one of the experiments we selected for reproduction was discovered to have flaws that made the meaningfulness of conducting a reproduction questionable. As a result, we had to change our coordinated study design from a reproduce approach to a standardise-then-reproduce-twice approach. Our overall (negative) finding that the great majority of human evaluations in NLP is not repeatable and/or not reproducible and/or too flawed to justify reproduction, paints a dire picture, but presents an opportunity for a rethink about how to design and report human evaluations in NLP.

Original language	English
Title of host publication	The Fourth Workshop on Insights from Negative Results in NLP
Place of Publication	Dubrovnik, Croatia
Publisher	Association for Computational Linguistics
Pages	1-10
Number of pages	10
Publication status	Published - 1 May 2023

Access to Document

2023.insights-1.1Final published version, 184 KBLicence: CC BY

https://aclanthology.org/2023.insights-1.1Licence: CC BY

Cite this

Belz, A., Thomson, C., Reiter, E., Abercrombie, G., Alonso-Moral, J. M., Arvan, M., Cheung, J., Cieliebak, M., Clark, E., Deemter, K. V., Dinkar, T., Dušek, O., Eger, S., Fang, Q., Gatt, A., Gkatzia, D., González-Corbelle, J., Hovy, D., Hürlimann, M., ... Yang, D. (2023). Missing Information, Unresponsive Authors, Experimental Flaws: The Impossibility of Assessing the Reproducibility of Previous Human Evaluations in NLP. In The Fourth Workshop on Insights from Negative Results in NLP (pp. 1-10). Association for Computational Linguistics. https://aclanthology.org/2023.insights-1.1

@inproceedings{502e63dea81d4a81b083012ef2cf0dea,

title = "Missing Information, Unresponsive Authors, Experimental Flaws: The Impossibility of Assessing the Reproducibility of Previous Human Evaluations in NLP",

abstract = "We report our efforts in identifying a set of previous human evaluations in NLP that would be suitable for a coordinated study examining what makes human evaluations in NLP more/less reproducible. We present our results and findings, which include that just 13% of papers had (i) sufficiently low barriers to reproduction, and (ii) enough obtainable information, to be considered for reproduction, and that all but one of the experiments we selected for reproduction was discovered to have flaws that made the meaningfulness of conducting a reproduction questionable. As a result, we had to change our coordinated study design from a reproduce approach to a standardise-then-reproduce-twice approach. Our overall (negative) finding that the great majority of human evaluations in NLP is not repeatable and/or not reproducible and/or too flawed to justify reproduction, paints a dire picture, but presents an opportunity for a rethink about how to design and report human evaluations in NLP.",

author = "Anya Belz and Craig Thomson and Ehud Reiter and Gavin Abercrombie and Alonso-Moral, {Jose M.} and Mohammad Arvan and Jackie Cheung and Mark Cieliebak and Elizabeth Clark and Deemter, {Kees van} and Tanvi Dinkar and Ond{\v r}ej Du{\v s}ek and Steffen Eger and Qixiang Fang and Albert Gatt and Dimitra Gkatzia and Javier Gonz{\'a}lez-Corbelle and Dirk Hovy and Manuela H{\"u}rlimann and Takumi Ito and Kelleher, {John D.} and Filip Klubicka and Huiyuan Lai and Lee, {Chris van der} and Miltenburg, {Emiel van} and Yiru Li and Saad Mahamood and Margot Mieskes and Malvina Nissim and Natalie Parde and Ond{\v r}ej Pl{\'a}tek and Verena Rieser and Romero, {Pablo Mosteiro} and Joel Tetreault and Antonio Toral and Xiaojun Wan and Leo Wanner and Lewis Watson and Diyi Yang",

year = "2023",

month = may,

day = "1",

language = "English",

pages = "1--10",

booktitle = "The Fourth Workshop on Insights from Negative Results in NLP",

publisher = "Association for Computational Linguistics",

address = "United States",

}

Belz, A, Thomson, C, Reiter, E, Abercrombie, G, Alonso-Moral, JM, Arvan, M, Cheung, J, Cieliebak, M, Clark, E, Deemter, KV, Dinkar, T, Dušek, O, Eger, S, Fang, Q , Gatt, A, Gkatzia, D, González-Corbelle, J, Hovy, D, Hürlimann, M, Ito, T, Kelleher, JD, Klubicka, F, Lai, H, Lee, CVD, Miltenburg, EV, Li, Y, Mahamood, S, Mieskes, M, Nissim, M, Parde, N, Plátek, O, Rieser, V, Romero, PM, Tetreault, J, Toral, A, Wan, X, Wanner, L, Watson, L & Yang, D 2023, Missing Information, Unresponsive Authors, Experimental Flaws: The Impossibility of Assessing the Reproducibility of Previous Human Evaluations in NLP. in The Fourth Workshop on Insights from Negative Results in NLP. Association for Computational Linguistics, Dubrovnik, Croatia, pp. 1-10. <https://aclanthology.org/2023.insights-1.1>

Missing Information, Unresponsive Authors, Experimental Flaws: The Impossibility of Assessing the Reproducibility of Previous Human Evaluations in NLP. / Belz, Anya; Thomson, Craig; Reiter, Ehud et al.
The Fourth Workshop on Insights from Negative Results in NLP. Dubrovnik, Croatia: Association for Computational Linguistics, 2023. p. 1-10.

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › Academic › peer-review

TY - GEN

T1 - Missing Information, Unresponsive Authors, Experimental Flaws: The Impossibility of Assessing the Reproducibility of Previous Human Evaluations in NLP

AU - Belz, Anya

AU - Thomson, Craig

AU - Reiter, Ehud

AU - Abercrombie, Gavin

AU - Alonso-Moral, Jose M.

AU - Arvan, Mohammad

AU - Cheung, Jackie

AU - Cieliebak, Mark

AU - Clark, Elizabeth

AU - Deemter, Kees van

AU - Dinkar, Tanvi

AU - Dušek, Ondřej

AU - Eger, Steffen

AU - Fang, Qixiang

AU - Gatt, Albert

AU - Gkatzia, Dimitra

AU - González-Corbelle, Javier

AU - Hovy, Dirk

AU - Hürlimann, Manuela

AU - Ito, Takumi

AU - Kelleher, John D.

AU - Klubicka, Filip

AU - Lai, Huiyuan

AU - Lee, Chris van der

AU - Miltenburg, Emiel van

AU - Li, Yiru

AU - Mahamood, Saad

AU - Mieskes, Margot

AU - Nissim, Malvina

AU - Parde, Natalie

AU - Plátek, Ondřej

AU - Rieser, Verena

AU - Romero, Pablo Mosteiro

AU - Tetreault, Joel

AU - Toral, Antonio

AU - Wan, Xiaojun

AU - Wanner, Leo

AU - Watson, Lewis

AU - Yang, Diyi

PY - 2023/5/1

Y1 - 2023/5/1

N2 - We report our efforts in identifying a set of previous human evaluations in NLP that would be suitable for a coordinated study examining what makes human evaluations in NLP more/less reproducible. We present our results and findings, which include that just 13% of papers had (i) sufficiently low barriers to reproduction, and (ii) enough obtainable information, to be considered for reproduction, and that all but one of the experiments we selected for reproduction was discovered to have flaws that made the meaningfulness of conducting a reproduction questionable. As a result, we had to change our coordinated study design from a reproduce approach to a standardise-then-reproduce-twice approach. Our overall (negative) finding that the great majority of human evaluations in NLP is not repeatable and/or not reproducible and/or too flawed to justify reproduction, paints a dire picture, but presents an opportunity for a rethink about how to design and report human evaluations in NLP.

AB - We report our efforts in identifying a set of previous human evaluations in NLP that would be suitable for a coordinated study examining what makes human evaluations in NLP more/less reproducible. We present our results and findings, which include that just 13% of papers had (i) sufficiently low barriers to reproduction, and (ii) enough obtainable information, to be considered for reproduction, and that all but one of the experiments we selected for reproduction was discovered to have flaws that made the meaningfulness of conducting a reproduction questionable. As a result, we had to change our coordinated study design from a reproduce approach to a standardise-then-reproduce-twice approach. Our overall (negative) finding that the great majority of human evaluations in NLP is not repeatable and/or not reproducible and/or too flawed to justify reproduction, paints a dire picture, but presents an opportunity for a rethink about how to design and report human evaluations in NLP.

M3 - Conference contribution

SP - 1

EP - 10

BT - The Fourth Workshop on Insights from Negative Results in NLP

PB - Association for Computational Linguistics

CY - Dubrovnik, Croatia

ER -

Missing Information, Unresponsive Authors, Experimental Flaws: The Impossibility of Assessing the Reproducibility of Previous Human Evaluations in NLP

Abstract

Access to Document

Fingerprint

Cite this