Empirical Evaluation of Public HateSpeech Datasets

TL;DR

Public hate-speech datasets are varied and often biased, posing challenges for training robust detectors. This work conducts a comprehensive empirical evaluation of ten public datasets, harmonising labels to a binary Hate/Not-Hate scheme, balancing data, and training a BERT-based baseline to compare mono-dataset and cross-dataset generalisation. Key findings show that content quality drives classifier performance more than size or modality, with Qian’s dataset supporting stronger cross-dataset transfer and Gomez’s data showing poor learnability; the study also links dataset features to performance via statistical analyses, offering actionable guidance for dataset curation. The results underscore the need for higher-quality, often multi-label, datasets and suggest concrete directions for improving hate-speech detection in real-world, cross-domain settings.

Abstract

Despite the extensive communication benefits offered by social media platforms, numerous challenges must be addressed to ensure user safety. One of the most significant risks faced by users on these platforms is targeted hate speech. Social media platforms are widely utilised for generating datasets employed in training and evaluating machine learning algorithms for hate speech detection. However, existing public datasets exhibit numerous limitations, hindering the effective training of these algorithms and leading to inaccurate hate speech classification. This study provides a comprehensive empirical evaluation of several public datasets commonly used in automated hate speech classification. Through rigorous analysis, we present compelling evidence highlighting the limitations of current hate speech datasets. Additionally, we conduct a range of statistical analyses to elucidate the strengths and weaknesses inherent in these datasets. This work aims to advance the development of more accurate and reliable machine learning models for hate speech detection by addressing the dataset limitations identified.

Figures (7)

• Figure 1: Block diagram of the T-test procedure used to compare the usage of hate terms in hateful and non-hateful speech in each dataset
• Figure 2: Baseline architecture
• Figure 3: Confusion matrix: mono-dataset classification error analyses
• Figure 4: Confusion matrix: generalised model train on Qian et al. dataset and tested on multiple dataset
• Figure 5: Confusion matrix: generalised classifier trained on Gomez et al. dataset and tested on multiple dataset