cs.MAMar 23, 2026
SaveIs AI Ready for Multimodal Hate Speech Detection? A Comprehensive Dataset and Benchmark Evaluation
Organizations: MOE KLINNS Lab, Xi’an Jiaotong University · The Hong Kong University of Science and Technology (Guangzhou) · School of Cyber Science and Engineering, Xi’an Jiaotong University · Northwest University
Abstract
Hate speech online targets individuals or groups based on identity attributes and spreads rapidly, posing serious social risks. Memes, which combine images and text, have emerged as a nuanced vehicle for disseminating hate speech, often relying on cultural knowledge for interpretation. However, existing multimodal hate speech datasets suffer from coarse-grained labeling and a lack of integration with surrounding discourse, leading to imprecise and incomplete assessments. To bridge this gap, we propose an agentic annotation framework that coordinates seven specialized agents to generate hierarchical labels and rationales. Based on this framework, we construct M^3 (Multi-platform, Multi-lingual, and Multimodal Meme), a dataset of 2,455 memes collected from X, 4chan, and Weibo, featuring fine-grained hate labels and human-verified rationales. Benchmarking state-of-the-art Multimodal Large Language Models reveals that these models struggle to effectively utilize surrounding post context, which often fails to improve or even degrades detection performance. Our finding highlights the challenges these models face in reasoning over memes embedded in real-world discourse and underscores the need for a context-aware multimodal architecture. Our dataset and code are available at https://github.com/mira-ai-lab/M3.
Figures & tables
Figure 1: Comparison between existing datasets and ours (M 3 ). Existing datasets typically label the meme (right) as normal . However, our dataset labels it as hate with a refined classification of healthy state -because of its accompanying post.
| Dataset | Domain | Label | Img text | Post | Rationale | Method | Source | |
| Hateful? | Categories | |||||||
| The Hateful Memes Challenge Set | Multiple fields | Hateful, Not-hateful | - | - | - | Human-only | Synthetical | |
| HatReD | Multiple fields | Hateful, Not-hateful | - | - | Human-only | Synthetical | ||
| HarMeme | COVID-19 | Very harmful, Partially harmful, Harmless | - | - | - | Human-only | Google Image Search, Reddit, Facebook, Instagram | |
| MMHS150K | Multiple fields | - | Not hate, Racist, Sexist, Homophobic, Religion, Other hate | - | - | Human-only | ||
| MAMI | Misogyny | Misogynous, Not-misogynous | Shaming, Stereotype, Objectification, Violence | - | - | Human-only | Twitter, Reddit | |
Table 1: Comparison of representative multimodal hate speech datasets. While existing datasets often provide coarse labels or limited context, OURS introduces fine-grained multi-dimensional annotations and additionally includes the surrounding post content, enabling richer and more context-aware hate speech analysis.
Figure 2: The agentic annotation framework for M 3 . ❶ Acquisition: Collector harvests multi-platform images and metadata. ❷ Preprocessing: Extractor and Cleaner perform OCR and metadata refinement. ❸ Annotation: Annotators, Arbiter, and Explicator collaborate on classification and rationale generation. ❹ Validation: Validator conducts quality assurance to finalize the M 3 dataset (sample entry on the right).
Figure 3: Visualizing linguistic patterns in M 3 . The top panel displays the word cloud of posts in hate samples from X, while the bottom-left and bottom-right panels illustrate the word cloud of Weibo and 4chan, respectively.
Figure 4: Hierarchical categories in M 3 . Hate samples are categorized into eight themes, with 1,018 single-labeled and 300 multi-labeled samples.
Figure 5: The distribution of single rationale and multiple rationales across X, Weibo, and 4chan.
Table 2: Comparison of state-of-the-art MLLMs on the M 3 dataset. The models are listed in order from open-source to proprietary, following their chronological release or version evolution: (1) the early LLaVA-v1.6 series (7B, 13B), (2) the reasoning-enhanced GLM-4.1V-9B-Thinking, (3) the latest Qwen series (Qwen2.5-VL-3B/7B and Qwen3-VL-8B), and (4) the proprietary GPT-4o and Gemini-3. Results cover binary and multi-class classification, and rationale quality under meme-only and meme + post settings. Boldface denotes the extremal values under the meme-only setting, while boldface with underline denotes the extremal values under the meme + post setting.
Figure 6: Performance comparison of MLLMs across three platforms. From left to right: 4chan (English-centric), X (Multi-lingual, e.g., Latin and Arabic), and Weibo (Chinese-dominant).
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
| Field | Description |
| post_id | Unique identifier of the post |
| post_time | Timestamp the post was published |
| img_id | Unique identifier for the meme |
| img_url | URL linking to the meme |
| post_text | Cleaned textual content of the post |
| img_text | OCR-extracted textual content within the meme |
Table 3: Fields in JSON Metadata Files. post_text contains the textual content of the post after Extractor and Cleaner.
Figure 7: Web-based arbitration interface for binary hate annotation. From left to right, examples are sourced from 4chan, X, and Weibo. The interface displays the meme alongside predictions from annotators to support human review of low-consistency samples.
Figure 8: Web-based arbitration interface for hate category annotation. From left to right, examples are sourced from 4chan, X, and Weibo. The interface presents hate categories with suggestions of annotators, allowing human arbiters to select, combine, or override model predictions.
| Validator | Agreement (1) | Opposition (0) |
| Graduate 1 | 195 | 5 |
| Graduate 2 | 194 | 6 |
| Graduate 3 | 189 | 11 |
| Vote Statistics | ||
| Three votes in favor | 181 | |
| Two votes in favor | 16 | |
Table 4: Validator voting results for a random audit of 200 high-consistency samples. The table reports individual agreement/opposition counts and aggregated vote statistics.
| ID | Category | Post | Rationale |
| 405 | Religion | Everyone but Christians refuse to participate in usury. | Blame Christians while using apocalyptic language |
| 502 | Politics | Mana. I had such high hopes for her… Then she came out as a commie. | Incite anti-communist hatred |
| 653 | Race | If there’s one speck of… is it still just good ol’e milk? | Insult Black people |
| 714 | Gender | Trannies be like “it’s not a mental illness”… | Depreciate transgender individuals |
| 832 | Health Status | A big thing I used to do… but just for me these days. | Discriminate against people with depression |
| 930 | Violence | Can’t come soon enough | Encourage suicide |
Table 5: Representative single-label examples for each hate category.
| Statistics | Value |
| Post Text Language Distribution | |
| Total posts | 2455 |
| Monolingual posts | 2117 |
| * Monolingual – Latin | 1794 |
| * Monolingual – Chinese | 323 |
| Multilingual posts | 338 |
Table 6: Language statistics of post and image text.
| Meme + Post | ||||||||||
| Platform | Model | Acc | P | R | F1 | Macro-P | Macro-R | Macro-F1 | Subset Acc | BERTScore |
| 4chan | LLaVA-v1.6-Vicuna-13B-hf | 55 | 81.58 | 52.54 | 63.92 | 43.03 | 63.13 | 47.56 | 16.1 | 95.89 |
| GLM-4.1V-9B-Thinking | 72.93 | 91.8 | 70.62 | 79.83 | 53.92 | 82.73 | 63.79 | 29.66 | 97.76 | |
| Qwen3-VL-8B-Instruct | 82.43 | 87.71 | 89.36 | 88.53 | 54.56 | 78.66 | 62.06 | 29.28 | 97.41 | |
| Gemini-3 | 69.36 | 86.09 | 71.09 | 77.88 | 46.9 | 86.01 | 58.74 | 20.83 | 97.46 | |
| X | LLaVA-v1.6-Vicuna-13B-hf | 64.26 | 23.08 | 18.18 | 20.34 | 37.42 | 53.55 | 39.87 | 9.85 | 96.04 |
Table 7: Multi-platform and multi-lingual evaluation results under two input settings: Meme + Post and Meme Only .
Figure 9: Case study showing model predictions under meme-only and meme+post settings. Adding a post changes the prediction from normal to hate (top), and refines category predictions (middle). The bottom part compares model-generated rationales, where GLM-4.1V aligns better with ground truth than GPT-4o and Qwen2.5-VL.
| Dataset | Size | Label | Classification | Language | methods |
| Waseem’s | 16,914 | Sexist, Racist, Neither | Multi-class | English | Manual |
| Founta’s | 80,000 | Offensive, Abusive, Hate speech, Aggressive, Cyberbullying, Spam, Normal | Multi-class | English | Manual |
| L-HSAB | 5846 | hate, abusive, normal | Multi-class | Arabic | Manual |
| CHSD | 17,430 | hate, normal | binary | Chinese | Manual |
| OLID | 14,100 | offensive, not offensive | Hierarchical multi-label | English | manual |
| targeted insult, untargeted |
Table 8: Summary of unimodal hate speech datasets. For OLID, targeted insult and untargeted are secondary labels under the primary label offensive ; further, individual , group , and other are tertiary labels under targeted insult . For HateXplain, hate , offensive , and normal are primary labels, while secondary labels denote specific targeted communities, such as African , Islam , and others. For HATEDAY, hateful , offensive , and neutral are primary labels; politics , national origin , gender , religion , sexual orientation are secondary labels under hateful . For HateBRXplain, highly , moderately , slightly are secondary labels under the primary label offensive ; Tertiary labels such as xenophobia are hate speech groups.
More in this topic
Hate Speech Detection
Hate Speech Detection identifies hateful or abusive language, its targets and intent, to support safer and more accountable content moderation.
See all 97 papers