Research map / Architectures and efficiency / Attention and transformers
Aligning Attention with Human Rationales for Self-Explaining Hate Speech Detection
Brage Eilertsen, Røskva Bjørgfinsdóttir, Francielle Vargas, Ali Ramezani-Kebrya
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add attention alignment supervision (alpha 0 to 100) (ablation) | BERT-base-uncased · HateXplain test | IoU F1: 0.019 -> 0.572 | Improved |
| add attention alignment supervision (alpha 0 to 100) (ablation) | BERT-base-uncased · HateXplain test | Token F1: 0.122 -> 0.678 | Improved |
| increase alpha from 0 to 100 (ablation) | BERT-base-uncased · HateXplain test | Macro F1 | No clear effect |
| increase alpha from 0 to 100 (ablation) | BERT-base-uncased · HateXplain test | GMB-Subgroup AUC: range 0.7112-0.7220 | No clear effect |
| add attention alignment supervision (alpha 0.05 to 100) (ablation) | BERTimbau-base · HateBRXplain test | IoU F1: 0.387 -> 0.751; baseline 0 | Improved |
| add attention alignment supervision (ablation) | BERTimbau-base · HateBRXplain test | Macro F1: 0.903 -> 0.921 (alpha=0.5) | No clear effect |
| add attention alignment supervision (ablation) | BERTimbau-base · HateBRXplain test | GMB-Subgroup AUC: 0.7165 -> 0.7157 | No clear effect |
| add attention alignment supervision (ablation) | BERT-base-uncased · HateXplain test | attention-rationale Pearson correlation: -0.084 -> 0.649 | Improved |
| add attention alignment supervision on identity terms (ablation) | BERT-base-uncased · HateXplain test | token precision (identity terms): 0.265 -> 0.938 | Improved |
| use SRA instead of BERT-MRP | BERT-base-uncased · HateXplain test | Macro F1: 0.699 -> 0.682 | Hurt |
| use SRA instead of BERT-MRP | BERT-base-uncased · HateXplain test | GMB-Subgroup AUC: 0.815 -> 0.714 | Hurt |
| use SRA instead of BERT-HateXplain supervised attention | BERT-base-uncased · HateXplain test | IoU F1: 0.120 -> 0.539 | Improved |
Closest papers by meaning
- Spanning the Spectrum of Hatred Detection: A Persian Multi-Label Hate Speech Dataset with Annotator Rationales (AAAI 2024 · similarity 0.87)
- Multi-Modal Discussion Transformer: Integrating Text, Images and Graph Transformers to Detect Hate Speech on Social Media (AAAI 2024 · similarity 0.86)
- TRACE: Textual Relevance Augmentation and Contextual Encoding for Multimodal Hate Detection (AAAI 2026 · similarity 0.85)
- A Multi-Task Benchmark for Abusive Language Detection in Low-Resource Settings (NeurIPS 2025 · similarity 0.84)
- IndicCONAN: A Multilingual Dataset for Combating Hate Speech in Indian Context (AAAI 2024 · similarity 0.83)
- Hateful Meme Detection through Context-Sensitive Prompting and Fine-Grained Labeling (Student Abstract) (AAAI 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
