Human evaluation is crucial for NLG systems as it provides a reliable assessment of the quality, effectiveness, and utility of generated language outputs. However, concerns about the reproducibility of such evaluations have emerged, casting doubt on the reliability and generalisability of reported results. In this paper, we present the findings of a reproducibility study on a data-to-text system, conducted under two conditions: (1) replicating the original setup as closely as possible with evaluators from AMT, and (2) replicating the original human evaluation but this time, utilising evaluators with a background in academia. Our experiments show that there is a loss of statistical significance between the original and reproduction studies, i.e. the human evaluation results are not reproducible. In addition, we found that employing local participants led to more robust results. We finally discuss lessons learned, addressing the challenges and best practices for ensuring reproducibility in NLG human evaluations.

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.