Features or Spurious Artifacts? Data-centric Baselines for Fair and Robust Hate Speech Detection

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3027 - 3040 July 10-15, 2022 ©2022 Association for Computational Linguistics Features or Spurious Artifacts? Data-centric Baselines for Fair and Robust Hate Speech Detection Alan Ramponi Fondazione Bruno Kessler Trento, Italy alramponi@fbk.eu Sara Tonelli Fondazione Bruno Kessler Trento, Italy satonelli@fbk.eu Abstract � Warning: this paper contains content that may be offensive or upsetting. Avoiding to rely on dataset artifacts to pre- dict hate speech is at the cornerstone of ro- bust and fair hate speech detection. In this pa- per we critically analyze lexical biases in hate speech detection via a cross-platform study, disentangling various types of spurious and authentic artifacts and analyzing their impact on out-of-distribution fairness and robustness. We experiment with existing approaches and propose simple yet surprisingly effective data- centric baselines. Our results on English data across four platforms show that distinct spuri- ous artifacts require different treatments to ul- timately attain both robustness and fairness in hate speech detection. To encourage research in this direction, we release all baseline mod- els and the code to compute artifacts, pointing it out as a complementary and necessary addi- tion to the data statements practice.1 1 Introduction Hate speech in online social communities is a se- rious and pervasive concern, which requires fair and robust automated approaches to be tackled at scale. However, despite the great progress in natu- ral language processing for detecting hate speech, current models have shown to be brittle when ap- plied to real-world data, exhibiting limited out-of- distribution (OOD) robustness (Vidgen et al., 2019) and perpetuating and amplifying harmful social bi- ases (Röttger et al., 2021). Noticeably, hate speech detection systems are typically trained on data from limited language varieties such as individual plat- forms, which inevitably exhibit differences in writ- ing norms, language use, and hate targets, hamper- ing generalization (Vidgen and Derczynski, 2020). One of the main reasons for limited robustness and fairness of mainstream hate speech detection 1Code and resources are available at https://github. com/dhfbk/hate-speech-artifacts. fair robust what have jews done to you?  RT [user]: I’m mad at this  All black people literally go there   Table 1: Posts wrongly labeled as hateful by a fine- tuned BERT classifier due to the presence of spurious lexical artifacts (identity and non identity-related) and their negative impact () on fairness and robustness. systems is largely ascribable to spurious statisti- cal correlations between surface lexical items and labels in training data, which models exploit to derive predictions. These biases are commonly referred to as lexical dataset artifacts, and have recently attracted attention in the NLP community, particularly in natural language inference (NLI) studies (Belinkov et al., 2019; Gururangan et al., 2018; Poliak et al., 2018, inter alia). Efforts to tackle the issue in hate speech detection are instead rather scattered, and mainly focus on fairness us- ing datasets from few platforms (Zhou et al., 2021; Kennedy et al., 2020b, inter alia), …

Content truncated to 500 words