# MMBERT: Scaled Mixture-of-Experts Multimodal BERT for Robust Chinese Hate Speech Detection under Cloaking Perturbations

arXiv: https://arxiv.org/abs/2508.00760 · HTML: https://mlsanity.com/paper/2508.00760/

## Key facts

- Cited by 2 papers in the map (data through 14 Oct 2026).
- Placed in Multimodal MoE models, one of 12 directions in Mixture-of-experts.
- Builds on 19 of its 35 references that are in the map.
- First posted on arXiv 1 Aug 2025.

## Placed in

- Mixture-of-experts / [Multimodal MoE models](https://mlsanity.com/moe/multimodal-moe-models/)

## Builds on (most cited first, 19 of 19)

- [Language Models are Few-Shot Learners](https://mlsanity.com/paper/2005.14165/) (2020)
- [RoBERTa: A Robustly Optimized BERT Pretraining Approach](https://mlsanity.com/paper/1907.11692/) (2019)
- [Robust Speech Recognition via Large-Scale Weak Supervision](https://mlsanity.com/paper/2212.04356/) (2022)
- [Automated Hate Speech Detection and the Problem of Offensive Language](https://mlsanity.com/paper/1703.04009/) (2017)
- [The Dawn of LMMs: Preliminary Explorations with GPT-4V(ision)](https://mlsanity.com/paper/2309.17421/) (2023)
- [ERNIE: Enhanced Representation through Knowledge Integration](https://mlsanity.com/paper/1904.09223/) (2019)
- [Learning Factored Representations in a Deep Mixture of Experts](https://mlsanity.com/paper/1312.4314/) (2013)
- [Racial Bias in Hate Speech and Abusive Language Detection Datasets](https://mlsanity.com/paper/1905.12516/) (2019)
- [SpeechT5: Unified-Modal Encoder-Decoder Pre-Training for Spoken Language Processing](https://mlsanity.com/paper/2110.07205/) (2021)
- [MoEfication: Transformer Feed-forward Layers are Mixtures of Experts](https://mlsanity.com/paper/2110.01786/) (2021)
- [Chinese CLIP: Contrastive Vision-Language Pretraining in Chinese](https://mlsanity.com/paper/2211.01335/) (2022)
- [COLD: A Benchmark for Chinese Offensive Language Detection](https://mlsanity.com/paper/2201.06025/) (2022)
- [ChineseBERT: Chinese Pretraining Enhanced by Glyph and Pinyin Information](https://mlsanity.com/paper/2106.16038/) (2021)
- [ERNIE-M: Enhanced Multilingual Representation by Aligning Cross-lingual Semantics with Monolingual Corpora](https://mlsanity.com/paper/2012.15674/) (2020)
- [CuMo: Scaling Multimodal LLM with Co-Upcycled Mixture-of-Experts](https://mlsanity.com/paper/2405.05949/) (2024)
- [Facilitating Fine-grained Detection of Chinese Toxic Language: Hierarchical Taxonomy, Resources, and Benchmarks](https://mlsanity.com/paper/2305.04446/) (2023)
- [ToxiCloakCN: Evaluating Robustness of Offensive Language Detection in Chinese with Cloaking Perturbations](https://mlsanity.com/paper/2406.12223/) (2024)
- [Can Prompting LLMs Unlock Hate Speech Detection across Languages? A Zero-shot and Few-shot Study](https://mlsanity.com/paper/2505.06149/) (2025)
- [GlyphCRM: Bidirectional Encoder Representation for Chinese Character with its Glyph](https://mlsanity.com/paper/2107.00395/) (2021)

## Built on by (most cited first, 2 of 2)

- [STEMTOX: From Collaborative Tags to Fine-Grained Toxic Meme Detection via Entropy-Guided Multi-Task Learning](https://mlsanity.com/paper/2508.04166/) (2025)
- [ToxiTrace: Gradient-Aligned Training for Explainable Chinese Toxicity Detection](https://mlsanity.com/paper/2604.12321/) (2026)

## Abstract (arXiv)

Hate speech detection on Chinese social networks presents distinct challenges, particularly due to the widespread use of cloaking techniques designed to evade conventional text-based detection systems. Although large language models (LLMs) have recently improved hate speech detection capabilities, the majority of existing work has concentrated on English datasets, with limited attention given to multimodal strategies in the Chinese context. In this study, we propose MMBERT, a novel BERT-based multimodal framework that integrates textual, speech, and visual modalities through a Mixture-of-Experts (MoE) architecture. To address the instability associated with directly integrating MoE into BERT-based models, we develop a progressive three-stage training paradigm. MMBERT incorporates modality-specific experts, a shared self-attention mechanism, and a router-based expert allocation strategy to enhance robustness against adversarial perturbations. Empirical results in several Chinese hate speech datasets show that MMBERT significantly surpasses fine-tuned BERT-based encoder models, fine-tuned LLMs, and LLMs utilizing in-context learning approaches.