Research map / Vision-language and multimodal / Cross-modal learning
MMBERT: Scaled Mixture-of-Experts Multimodal BERT for Robust Chinese Hate Speech Detection Under Cloaking Perturbations
Qiyao Xue, Yuchen Dou, Zheyuan Ryan Shi, Xiang Lorraine Li, Wei Gao
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 speech modality to text-only MMBERT (ablation) | MMBERT · ToxiCloakCN | macro F1: 86.1 -> 91.1 | Improved |
| add vision modality to text-only MMBERT (ablation) | MMBERT · ToxiCloakCN | macro F1: 86.1 -> 86.6 | No clear effect |
| add speech modality to text-only MMBERT (ablation) | MMBERT · COLD | macro F1: 81.5 -> 83.8 | Improved |
| add both speech and vision with MoE and 3-stage training (full MMBERT vs text-speech) (ablation) | MMBERT · ToxiCloakCN | macro F1: 91.1 -> 95.2 | Improved |
| add progressive three-stage training (aligner and expert stages) (ablation) | MMBERT · ToxiCloakCN | validation accuracy / training loss | Improved |
| use MMBERT instead of ChineseBERT | fine-tuned encoders · ToxiCloakCN | macro F1: 86.8 -> 95.2 | Improved |
| use MMBERT instead of ChineseBERT | fine-tuned encoders · ToxiCN | macro F1: 90.6 -> 92.2 | Improved |
| add perturbed examples and CoT to GPT-4o prompt | GPT-4o · ToxiCloakCN | macro F1: 62.4 -> 69.3 | Improved |
| use CoT prompting and perturbed examples | DeepSeek-v3 · ToxiCN | macro F1: 71.7 -> 77.1 (basic to CoT) while ToxiCloakCN 66.2 -> 71.6 | No clear effect |
| fine-tune LLaMA3-8B vs prompting LLaMA3-8B (2 examples) | LLaMA3-8B · ToxiCloakCN | macro F1: 68.0 -> 79.3 | Improved |
Later papers that compared with this method
1 later accepted paper reports MMBERT as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Emotion and Intention Guided Multi-Modal Learning for Sticker Response Selection (AAAI 2026 · better on 4 of 4 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- TRACE: Textual Relevance Augmentation and Contextual Encoding for Multimodal Hate Detection (AAAI 2026 · similarity 0.84)
- Multi-Modal Discussion Transformer: Integrating Text, Images and Graph Transformers to Detect Hate Speech on Social Media (AAAI 2024 · similarity 0.84)
- Spanning the Spectrum of Hatred Detection: A Persian Multi-Label Hate Speech Dataset with Annotator Rationales (AAAI 2024 · similarity 0.83)
- MLLMGuard: A Multi-dimensional Safety Evaluation Suite for Multimodal Large Language Models (NeurIPS 2024 · similarity 0.81)
- Bad Actor, Good Advisor: Exploring the Role of Large Language Models in Fake News Detection (AAAI 2024 · similarity 0.81)
- MSE-Adapter: A Lightweight Plugin Endowing LLMs with the Capability to Perform Multimodal Sentiment Analysis and Emotion Recognition (AAAI 2025 · similarity 0.81)
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.
