PaperlayerView exact revision on arXiv

Source-linked AI summary

ETHOS: an Online Hate Speech Detection Dataset

Ioannis Mollas, Zoe Chrysopoulou, Stamatis Karlos, Grigorios Tsoumakas

arXiv:2006.08328v2cs.CLcs.LGstat.ML

TL;DR

Online hate-speech detection requires reliable data despite noisy social-media text, imbalanced collections, and the ethical risks of erroneous classification. ETHOS constructs binary and multi-label datasets from YouTube and Reddit comments using active sampling and crowdsourced validation. The resulting dataset balances labels, supports multi-dimensional hate-speech analysis, and provides a basis for baseline experiments and reuse, while remaining constrained by limited data and human annotation uncertainty.

  • Problem

    Reliable hate-speech detection is limited by noisy social-media text, imbalanced datasets, and the need to represent overlapping hate categories rather than only binary outcomes.

  • Method

    ETHOS uses active sampling, structured annotation, and Figure-Eight crowdsourcing to construct binary and multi-label datasets from Hatebusters and Reddit comments.

  • Results

    ETHOS achieves balanced labels, including almost perfect balance across six hate-speech categories, and supports satisfactory learning performance across evaluated models.

  • Takeaways & Limitations

    ETHOS provides a reusable seed dataset and multi-label information for studying hate speech beyond the binary classification setting.

  • Takeaways & Limitations

    The dataset remains limited by crowdsourcing constraints, restricted budgets and traffic, human-intensive annotation, and annotation uncertainty.

Abstract

from arXiv · show

Online hate speech is a recent problem in our society that is rising at a steady pace by leveraging the vulnerabilities of the corresponding regimes that characterise most social media platforms. This phenomenon is primarily fostered by offensive comments, either during user interaction or in the form of a posted multimedia context. Nowadays, giant corporations own platforms where millions of users log in every day, and protection from exposure to similar phenomena appears to be necessary in order to comply with the corresponding legislation and maintain a high level of service quality. A robust and reliable system for detecting and preventing the uploading of relevant content will have a significant impact on our digitally interconnected society. Several aspects of our daily lives are undeniably linked to our social profiles, making us vulnerable to abusive behaviours. As a result, the lack of accurate hate speech detection mechanisms would severely degrade the overall user experience, although its erroneous operation would pose many ethical concerns. In this paper, we present 'ETHOS', a textual dataset with two variants: binary and multi-label, based on YouTube and Reddit comments validated using the Figure-Eight crowdsourcing platform. Furthermore, we present the annotation protocol used to create this dataset: an active sampling procedure for balancing our data in relation to the various aspects defined. Our key assumption is that, even gaining a small amount of labelled data from such a time-consuming process, we can guarantee hate speech occurrences in the examined material.

1 Introduction

Online hate speech is difficult to detect reliably because social-media content is noisy, data collection can be imbalanced, and comments may express multiple hate-related categories. ETHOS addresses these issues with a step-by-step, actively sampled annotation protocol and binary and multi-label dataset views.

  • Social-media anonymity and avoidance of legislation have contributed to the growth of hate-speech phenomena, prompting platforms to use AI for detection and prevention.
  • Short-text sparseness and noise, together with scalability problems in unfiltered collections, make fully automated hate-speech detection difficult.
  • Existing hate-speech datasets often suffer from imbalanced classes or redundant information because sampling stages were not carefully designed.
  • Multi-label classification captures overlapping hate categories and label dependencies that binary classification overlooks.
  • ETHOS presents a step-by-step process for creating a multi-labelled dataset intended to reduce common social-media data-collection problems and yield more informative instances.
  • The paper validates the dataset through crowdsourcing and evaluates binary and multi-label classification with traditional, ensemble, and neural models.

2 Related Datasets

Prior hate-speech datasets span platforms, languages, labels, and annotation settings, but their construction often involves targeted collection or skewed and redundant data. ETHOS is positioned as a protocol-based alternative built from social-media sources.

  • ETHOS uses Hatebusters data as seed data and introduces a protocol intended to overcome imbalanced or biased labels through preprocessing and structured dataset construction.
  • Existing datasets include Twitter, YouTube, forums, news comments, and multilingual social-media content, with varied labeling schemes and collection procedures.
  • Related datasets address binary hate-speech detection, offensive speech, misogyny, cyberbullying, sentence-level hate, and multilingual expressions.
  • The Hatebusters collection contained 454 YouTube comments, but semi-supervised expansion produced a more redundant variant of the original dataset.
  • Figure 2 presents the stages used to create the ETHOS dataset.

3 ETHOS Dataset Creation

ETHOS is created through a three-stage protocol combining platform-based collection, model-assisted prediction, active sampling, and crowdsourced validation. The resulting binary and multi-label datasets are designed to provide balanced, informative hate-speech annotations.

  • Data Collection: ETHOS combines Hatebusters and Reddit sources to collect comments for annotation.Hatebusters supplies YouTube-derived data, while Reddit comments are selected from archived subreddit files.
  • Data Prediction: A prediction stage trains models on labeled data and uses the best 10-fold-CV algorithm to assign labels to newly collected comments.The examined models include SVMs, Random Forests, Logistic Regression, and neural networks with word- or character-level TF-IDF features.
  • Data Annotation: Active annotation combines uncertainty sampling and maximum relevance to select informative comments for manual labeling until the target threshold T = 1.000 is reached.This procedure is intended to reduce redundancy while preserving label informativeness and limiting overfitting.
  • Data Validation: Figure-Eight validation asks contributors to identify hate speech and, when applicable, annotate violence and hate-speech categories.The category labels cover gender, race, national origin, disability, religion, and sexual orientation.
  • Dataset Overview: The resulting ETHOS files contain 998 comments with an isHate label and 433 hate-speech messages annotated across eight multi-label categories.The binary file records hate-speech presence or absence, while the multi-label file includes violence, directedness, and six target categories.
  • Dataset Configuration: ETHOS maintains balanced isHate classes and near-balanced hate-speech-category labels, while crowdsourcing remains subject to human uncertainty.The authors suggest more representative validation questions and interactive feedback as possible improvements.

4 Dataset Baseline Evaluation

ETHOS establishes binary and multi-label baseline evaluations using conventional machine-learning models and neural architectures, with preprocessing, vectorization, and task-specific metrics. Neural approaches generally outperform conventional methods, while NNBR achieves the strongest multi-label performance across nearly all reported metrics.

  • Evaluation setup: The evaluation preprocesses text with lowercasing, contraction transformation, punctuation removal, stemming, and lemmatization before vectorization.TF-IDF and Text-to-Sequences representations were examined, alongside FastText, GloVe, and BERT embeddings.
  • Binary classification: Binary experiments compare Naive Bayes, logistic regression, SVMs, random forests, gradient boosting, and several neural architectures, including CNN-, LSTM-, and BERT-based models.The neural models include attention-based CNNs, LSTMs, hybrid architectures, BiLSTMs, BERT, and DistilBERT.
  • Binary classification: Neural networks outperform conventional machine-learning techniques in the binary evaluation, while Bayesian learners perform lowest on almost every metric and tree ensembles trail SVMs and logistic regression.The evaluation reports accuracy, macro precision, recall, F1-score, confusion matrices, specificity, and sensitivity.
  • Binary classification: DistilBERT and BERT outperform the other binary models on every reported metric, with DistilBERT slightly exceeding BERT.Among embedding-based neural models, BiLSTMs achieve the strongest performance, and BiLSTMs with FastText and GloVe attain the highest hate-category recall and high accuracy.
  • Multi-label classification: For multi-label classification, the study evaluates problem-transformation and neural approaches using Hamming loss, subset accuracy, instance-based metrics, and label-based macro and micro measures.NNBR combines BiLSTMs, attention, feed-forward layers, and eight output units in a binary-relevance design; it achieves the highest score in 12 out of 13 metrics.

5 Dataset Experimentation

Experiments evaluate ETHOS through balanced-versus-random binary training, cross-dataset transfer, and multi-label generalisation. Results indicate that balanced annotation improves minority-class identification, while dataset mismatches constrain transfer across collections and labels.

  • 5.1 Balanced vs Random Comparison: Balanced ETHOS subsets improve SVM performance on held-out data, although random subsets can achieve higher accuracy on imbalanced D1.The authors attribute this contrast to majority-class bias in random training data and caution that weighted F1 comparisons are uncertain because DBa has high variance.
  • 5.1 Balanced vs Random Comparison: Balanced training better predicts the minority Hate class, whereas random training favors the majority No Hate class.The authors therefore regard balanced data as more valuable for identifying hateful comments despite its limited cardinality.
  • 5.2 Generalising on binary level: ETHOS and the 24-times-larger D1 dataset show similar SVM predictive ability in cross-dataset evaluation.The authors attribute this result to ETHOS’s more sophisticated data collection and annotation, which compensates for its smaller size.
  • 5.2 Generalising on binary level: Cross-dataset models do not predict each other’s hate instances reliably, partly because D1 contains uneven category distributions and potentially non-hateful offensive examples.Applying ETHOS multi-label predictions to D1 identified 326 race, 257 sexuality, and 230 gender instances among 1,430 hate tweets.
  • 5.2 Generalising on binary level: Combining ETHOS with D1 yields 55.27% balanced accuracy and 90.88% weighted F1, with 18.84% F1 for Hate Speech.The authors report improved overall learning behaviour when datasets with different dynamics are combined.
  • 5.3 Generalising on multi-label level: BiLSTM transfer from ETHOS to D2 predicts sexual_orientation perfectly and performs poorly on directed_vs_generalized, violence, race, and religion.Religion reaches 97.82% for negative instances but 27.31% for positive instances, while violence reaches 29.09% for positive cases.
  • 5.3 Generalising on multi-label level: D2’s missing race information limits evaluation, while mismatches between datasets produce unseen examples and erroneous race assignments.ETHOS contains no examples with the phrase “ching chong”; the model misclassified 35 non-race-labelled instances as race, 26 of which contained race-targeting hate speech.

6 Discussion

ETHOS is presented as a reusable hate-speech dataset and annotation protocol built around balancing fine-grained classes and actively selecting informative instances. Experiments support its learning capacity, while the authors identify data scarcity, social-media scope, and alternative sampling as important boundaries and directions.

  • Contributions: ETHOS provides a publicly reusable hate-speech dataset constructed from social-media data with balanced fine-grained classes.The balancing process uses a fine-tuned learner and active learning to reduce redundant information and annotation effort.
  • Method: The protocol emphasizes mining informative instances for both binary and multi-label classification while retaining discriminative hate-speech examples.This addresses the need to match annotation effort with instances that are useful for distinguishing hate-speech cases.
  • Results: Experiments found satisfactory learning performance across models, especially embedding-based models, and demonstrated ETHOS’s potential as a seed dataset for related datasets.The authors describe this as a proof of concept for exploiting the dataset’s learning capacity.
  • Limitations: Data scarcity remains a central limitation because crowdsourcing constraints and daily active instance selection make annotation human-intensive and costly.The authors mention restricted budgets, user traffic, and the continuing effort required to maintain a balanced target dataset.
  • Future directions: ETHOS can be combined with similar datasets, while hybrid weakly supervised approaches could merge semi-supervised and active learning to reduce human intervention.These are presented as directions for extending the dataset and its learning framework.
  • Scope and future directions: Because ETHOS contains relatively short social-media comments, models trained on it may fail on larger documents without segmentation.The authors also state that the general protocol could extend to larger databases and that alternative multi-label sampling strategies may reduce human effort.
Loading 2006.08328v2…