[论文解读] ETHOS: an Online Hate Speech Detection Dataset

Ioannis Mollas, Zoe Chrysopoulou+2 more|arXiv (Cornell University)|Jun 11, 2020

Hate Speech and Cyberbullying Detection参考文献 65被引用 57

一句话总结

ETHOS 提供一个小型、均衡的二分类与多标签仇恨言论数据集，通过主动学习标注协议和众包创建，基线结果显示 DistilBERT/BERT 表现最好。

ABSTRACT

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.

研究动机与目标

• Address the scarcity and imbalance in hate speech datasets by introducing a carefully annotated, balanced HS dataset.
• Develop a practical annotation protocol combining active learning and crowdsourcing to create reliable labels.
• Provide binary and multi-label HS datasets (Ethos_Binary.csv and Ethos_Multi_Label.csv) and baseline model evaluations.
• Assess inter-annotator reliability and dataset quality to support robust ML/HM research.
• Demonstrate the usefulness of the dataset for analyzing HS with multi-label semantics and label dependencies.

提出的方法

• Proposes a three-stage data collection/annotation protocol: Platform Selection & Data Collection, Data Prediction, Manual Data Annotation.
• Data collected from Hatebusters and Reddit, with an initial pre-trained SVM hatescore guiding sampling.
• Data prediction uses grid-tuned classifiers on expanded labelled data to select unlabelled comments for annotation.
• Manual annotation utilizes active learning uncertainty sampling and maximum relevance strategies to balance informativeness and label diversity.
• Validation via Figure-Eight crowdsourcing with five annotators per item and measurement of Fleiss’ kappa for reliability.
• Dataset configuration yields two files: Ethos_Binary.csv (998 comments, isHate) and Ethos_Multi_Label.csv (433 comments, 8 HS categories plus binary violence/directed_vs_generalized indicators).

实验结果

研究问题

• RQ1 How can HS datasets be created that are both balanced across classes and informative for ML models?
• RQ2 What is the performance of a wide range of models (traditional ML and neural nets) on binary HS detection using ETHOS?
• RQ3 How does multi-label HS annotation capture category dependencies and improve detection/interpretation compared to binary labeling?
• RQ4 What are the inter-annotator reliability measures when crowdsourcing HS labels on a carefully sampled dataset?
• RQ5 How do embedding-based models (FT, GV, BERT variants) compare in HS detection tasks on ETHOS?

主要发现

Model	F1 Score	F1 Hate	Accuracy	Precision	Sensitivity	Recall	Recall Hate	Specificity
MultinomialNB	63.78	59.14	64.73	64.06	58.82	63.96	59.45	69.2
BernoulliNB	47.78	44.52	48.3	48.23	47.81	48.16	41.65	48.51
Logistic Regression	66.5	64.35	66.94	66.94	68.78	67.07	60.46	65.36
SVM	66.07	63.77	66.43	66.47	68.08	66.7	59.96	65.32
Random Forests	64.41	60.07	65.04	64.69	60.61	64.68	59.54	68.75
Gradient Boosting	63.55	59.21	64.33	64.34	59.67	64.2	58.76	68.73
CNN+Attention+FT+GV	75.76	71.76	76.56	76.86	68.64	75.66	75.18	82.68
LSTM+FT+GV	75.24	72.24	75.95	76.57	72.11	75.53	72.36	78.95
FF+LSTM+CNN+FT+GV	75.49	72.08	76.15	76.29	70.88	75.52	73.28	80.16
BiLSTM+FT+GV	77.84	75.40	78.16	78.05	77.15	78.04	73.73	78.94
BERT	79.60	77.13	79.96	79.89	77.87	79.73	76.4	81.59
DistilBERT	79.92	77.16	80.36	80.28	76.47	79.91	77.87	83.36

• Neural models outperform traditional ML baselines on binary HS detection, with BiLSTM and CNN-based architectures achieving high F1 and recall.
• Fine-tuned BERT and DistilBERT yield the strongest overall performance, with DistilBERT slightly better in several metrics.
• BiLSTM+FT+GV achieves high recall for the hate category, indicating strong detection of HS instances.
• The dataset demonstrates balanced label distributions and substantial annotator agreement (Fleiss’ kappa values above 0.75 for HS-related labels).
• ETHOS provides a practical, active-learning driven protocol for creating balanced HS datasets, addressing common issues like redundancy and label bias.