Original title: Ethical and Socially Aware English–Norwegian Hate Speech Detection: A Systematic Design Framework
Ehtesham Hashmi, Sule Yildirim Yayilgan · September 2026
Hate speech targeting race, gender, religion, and sexual orientation spreads rapidly on social media. Detecting it fairly across multiple languages is hard. To solve this, researchers built an ethical framework for English and Norwegian text. They gathered 14K social media posts and labeled them for toxic language and targeted identities using human reviewers and helper models. Their system uses one model to flag whether a post is hateful, offensive, aggressive, or neutral, while also identifying targeted groups. To make it fair, they trained the system using balanced loss re-weighting and word-swapping techniques so simple identity words do not trigger false alarms. They also used an explainability tool called LIME to show which words drove each decision. Their NorBERT-small model achieved a 0.92 macro F1-score for toxicity, though some bias gaps remained between certain identity pairs.
Imagine you are in charge of a classroom and need to spot when someone is being mean about another student's background. You want to be fair and not punish someone for simply mentioning a topic. But you also must explain why you think a comment is mean. Now suppose the school gives you a big book of past comments to learn from, but some groups of students are barely mentioned in it. This paper builds a system that spots mean language online, tries to treat all groups fairly, and shows its reasoning, even when some groups have very few examples to learn from.
Social media platforms struggle to catch hate speech without unfairly flagging innocent posts. This work builds a detection system for English and Norwegian that tries to be fairer, not just more accurate. It shows how designers can bake ethical choices directly into the technology, not add them as an afterthought.
The results are specific to the collected dataset and may not generalize to other platforms, domains, or external datasets. Several social-attribute classes are sparsely represented, so aggregate F1-scores should be interpreted cautiously. Residual disparities remain for some identity pairs after mitigation. Cross-model comparisons are descriptive because model families use different architectures and training procedures.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Ethical and Socially Aware English–Norwegian Hate Speech Detection: A Systematic Design Framework — Ehtesham Hashmi, Sule Yildirim Yayilgan — Machine Learning and Knowledge Extraction — doi:10.3390/make8090292
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper.