Square One Bias in NLP: Towards a Multi-Dimensional Exploration of the Research Manifold

Findings of the Association for Computational Linguistics: ACL 2022, pages 2340 - 2354 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Square One Bias in NLP: Towards a Multi-Dimensional Exploration of the Research Manifold Sebastian Ruder∗ Google Research ruder@google.com Ivan Vuli´c∗ University of Cambridge iv250@cam.ac.uk Anders Søgaard∗ University of Copenhagen soegaard@di.ku.dk Abstract The prototypical NLP experiment trains a stan- dard architecture on labeled English data and optimizes for accuracy, without accounting for other dimensions such as fairness, inter- pretability, or computational efficiency. We show through a manual classification of recent NLP research papers that this is indeed the case and refer to it as the square one experi- mental setup. We observe that NLP research often goes beyond the square one setup, e.g, focusing not only on accuracy, but also on fairness or interpretability, but typically only along a single dimension. Most work tar- geting multilinguality, for example, considers only accuracy; most work on fairness or in- terpretability considers only English; and so on. Such one-dimensionality of most research means we are only exploring a fraction of the NLP research search space. We provide his- torical and recent examples of how the square one bias has led researchers to draw false conclusions or make unwise choices, point to promising yet unexplored directions on the re- search manifold, and make practical recom- mendations to enable more multi-dimensional research. We open-source the results of our an- notations to enable further analysis.1 1 Introduction Our categorization of objects, say screwdrivers or NLP experiments, is heavily biased by early pro- totypes (Sherman, 1985; Das-Smaal, 1990). If the first 10 screwdrivers we see are red and for hexagon socket screws, this will bias what features we learn to associate with screwdrivers. Likewise, if the first 10 NLP experiments we see or conduct are in sentiment analysis, this will likely also bias how we think of NLP experiments in the future. In this position paper, we postulate that we can meaningfully talk about the prototypical NLP ex- ∗The authors contributed equally to this work. 1github.com/google-research/url-nlp Figure 1: Visualization of contributions of ACL 2021 oral papers along 4 dimensions: multilinguality, fair- ness and bias, efficiency, and interpretability (indicated by color). Most work is clustered around the SQUARE ONE or along a single dimension. periment, and that the existence of such an exper- imental prototype steers and biases the research dynamics in our community. We will refer to this prototype as NLP’s SQUARE ONE—and to the bias that follows from it, as the SQUARE ONE BIAS. We argue this bias manifests in a particular way: Since research is a creative endeavor, and researchers aim to push the research horizon, most research papers in NLP go beyond this prototype, but only along a single dimension at a time. Such dimensions might include multilinguality, efficiency, fairness, and interpretability, among others. The effect of the SQUARE ONE BIAS is to baseline novel research contributions, rewarding work that differs from the prototype in a concise, one-dimensional way. We present several examples of …

Content truncated to 500 words