When building a predictive model, it is often difficult to ensure that application-specific requirements are encoded by the model that will eventually be deployed. Consider researchers working on hate speech detection. They will have an idea of what is considered hate speech, but building a model that reflects their view accurately requires preserving those ideals throughout the workflow of data set construction and model training. Complications such as sampling bias, annotation bias, and model misspecification almost always arise, possibly resulting in a gap between the application specification and the model’s actual behavior upon deployment. To address this issue for hate speech detection, we propose DefVerify: a 3-step procedure that (i) encodes a user-specified definition of hate speech, (ii) quantifies to what extent the model reflects the intended definition, and (iii) tries to identify the point of failure in the workflow. We use DefVerify to find gaps between definition and model behavior when applied to six popular hate speech benchmark datasets.

The subjectivity of automatic hate speech detection makes it a complex task, reflected in different and incomplete definitions in NLP. We present hate speech criteria, developed with insights from a law and social science expert, that help researchers create more explicit definitions and annotation guidelines on five aspects: (1) target groups and (2) dominance, (3) perpetrator characteristics, (4) explicit presence of negative interactions, and the (5) type of consequences/effects. Definitions can be structured so that they cover a more broad or more narrow phenomenon and conscious choices can be made on specifying criteria or leaving them open. We argue that the goal and exact task developers have in mind should determine how the scope of hate speech is defined. We provide an overview of the properties of datasets from hatespeechdata.com that may help select the most suitable dataset for a specific scenario.

Despite their success, modern language models are fragile. Even small changes in their training pipeline can lead to unexpected results. We study this phenomenon by examining the robustness of ALBERT (Lan et al., 2020) in combination with Stochastic Weight Averaging (SWA)—a cheap way of ensembling—on a sentiment analysis task (SST-2). In particular, we analyze SWA’s stability via CheckList criteria (Ribeiro et al., 2020), examining the agreement on errors made by models differing only in their random seed. We hypothesize that SWA is more stable because it ensembles model snapshots taken along the gradient descent trajectory. We quantify stability by comparing the models’ mistakes with Fleiss’ Kappa (Fleiss, 1971) and overlap ratio scores. We find that SWA reduces error rates in general; yet the models still suffer from their own distinct biases (according to CheckList).