Toxic, Hateful, Offensive or Abusive? What Are We Really Classifying? An Empirical Analysis of Hate Speech Datasets

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 6786–6794 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 6786 Toxic, Hateful, Offensive or Abusive? What Are We Really Classifying? An Empirical Analysis of Hate Speech Datasets Paula Fortuna1, Juan Soler-Company1, Leo Wanner2,1 1NLP Group, Pompeu Fabra University, 2Catalan Institute for Research and Advanced Studies (ICREA) paulatfortuna@gmail.com juan.soler|leo.wanner@upf.edu Abstract The field of the automatic detection of hate speech and related concepts has raised a lot of interest in the last years. Different datasets were annotated and classified by means of applying different machine learning algorithms. However, few efforts were done in order to clarify the applied categories and homogenize different datasets. Our study takes up this demand. We analyze six different publicly available datasets in this field with respect to their similarity and compatibility. We conduct two different experiments. First, we try to make the datasets compatible and represent the dataset classes as Fast Text word vectors analyzing the similarity between different classes in a intra and inter dataset manner. Second, we submit the chosen datasets to the Perspective API Toxicity classifier, achieving different performances depending on the categories and datasets. One of the main conclusions of these experiments is that many different definitions are being used for equivalent concepts, which makes most of the publicly available datasets incompatible. Grounded in our analysis, we provide guidelines for future dataset collection and annotation. Keywords: hate speech, toxicity, aggression, offensive, dataset comparison 1. Introduction Over the past years, the amount of online offensive speech has been growing steadily. To address this problem, a sig- nificant number of scientific publications focused on two different (albeit related) tasks: (i) the compilation and an- notation of corpora and (ii) the automatic detection of dif- ferent types of offensive speech, among them, e.g., toxi- city, hate, abuse, using generic state-of-the-art (i.e., ma- chine learning-based) natural language processing tech- niques. However, critics on this duality are starting to emerge in the community. The main concern resides in the restricted generalization potential of the trained ma- chine learning (ML) models for classification of offensive speech. For instance, Swamy et al. (2019) find that by training on top of BERT models (Devlin et al., 2018), it is possible to obtain a language model that performs very competitively for different datasets, however, the general- ization of the model depends highly on the training data. The authors hypothesize that a model will generalize better if it is used on data that is more similar to the data used for training, and worse if the data contains more non-offensive samples. In general, the limited generalization potential is a problem that can have two different origins. Firstly, online offensive speech varies depending on the targeted groups (e.g., the terms used to express hate against a community in Africa are clearly different from the terms used against African Americans) and on the context (e.g., the hate against the LGBT community is likely to be worded differently …

Content truncated to 500 words