Hate speech vs. offensive language classification is an NLP task that distinguishes between targeted hate speech directed at protected groups and generally offensive or vulgar language that may be crude but does not target specific identity groups. This distinction is critical for content moderation because the two require very different responses.

Definitions

• Hate Speech: Language that attacks, dehumanizes, or incites violence against people based on protected characteristics — race, ethnicity, religion, gender, sexual orientation, disability, or national origin.
• Offensive Language: Language that is vulgar, profane, rude, or crude but does not target a specific identity group. Swear words, insults, or aggressive language can be offensive without being hate speech.

Why the Distinction Matters

• Legal: Hate speech may violate laws in many countries. Offensive language is generally protected speech.
• Platform Policy: Social media platforms ban hate speech but typically allow offensive language. Misclassifying offensive language as hate speech results in unfair censorship.
• Impact: Hate speech can cause psychological harm to targeted communities, reinforce discrimination, and incite real-world violence.

Classification Challenges

• Context Sensitivity: The same word can be hate speech in one context and friendly banter in another (in-group reclaimed slurs).
• Implicit Hate: Coded language, dog whistles, and indirect references can convey hate without explicit slurs.
• Annotator Bias: Different annotators have different thresholds for what constitutes hate speech, leading to noisy labels.
• False Positives: Over-aggressive classifiers disproportionately flag content from minority communities who discuss hate speech or reclaim language.

Technical Approaches

• Multi-Class Classification: Three-class model — hate speech, offensive, neither. Fine-tuned BERT/RoBERTa models achieve good performance.
• Target Identification: Detect the target of the language — if it targets a protected group, more likely hate speech.
• Context Windows: Include surrounding conversation for context-dependent classification.

Datasets: Davidson et al. (2017) hate speech vs. offensive language dataset, HateXplain with rationale annotations, Gab Hate Corpus, Civil Comments.

This classification task requires nuanced understanding of language, context, and social dynamics — automated systems should be used as tools to assist human moderators rather than as sole decision-makers.