Missing Information, Unresponsive Authors, Experimental Flaws: The Impossibility of Assessing the Reproducibility of Previous Human Evaluations in NLP
Anya Belz, Craig Thomson, E. Reiter, Gavin Abercrombie, J. Alonso-Moral, Mohammad Arvan, J. Cheung, Mark Cieliebak, Elizabeth Clark, K. V. Deemter, Tanvi Dinkar, Ondrej Dusek, Steffen Eger, Qixiang Fang, Albert Gatt, Dimitra Gkatzia, Javier Gonz’alez-Corbelle, Dirk Hovy, Manuela Hurlimann, Takumi Ito, John D. Kelleher, Filip Klubicka, Huiyuan Lai, Chris van der Lee, Emiel van Miltenburg, Yiru Li, Saad Mahamood, Margot Mieskes, M. Nissim, Natalie Parde, Ondvrej Pl’atek, Verena Rieser, Pablo Romero, Joel Tetreault, Antonio Toral, Xiao-Yi Wan, L. Wanner, Lewis J. Watson, Diyi Yang
Insights from Negative Results in NLP @ EACL
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.