---
license: cc-by-nc-sa-4.0
task_categories:
  - visual-question-answering
language:
  - zh
  - en
tags:
  - meme-understanding
  - multimodal
  - cultural-knowledge
  - benchmark
size_categories:
  - 1K<n<10K
---

# MemeBench: Diagnosing Cultural-Semantic Understanding in LVLMs through Memes

## Dataset Summary

MemeBench is a bilingual diagnostic benchmark for open-ended meme interpretation, evaluating large vision-language models' cultural-semantic understanding. It contains **1,253 expert-annotated memes** spanning 7 cultural domains in Chinese and English.

Each meme is annotated with a structured **VIKR schema** covering four diagnostic layers:
- **V**isual clues: Can the model describe what it sees?
- **I**dentity links: Can the model identify who/what is depicted?
- **K**nowledge units: Does the model know the cultural background?
- **R**easoning mechanisms: Can the model explain why the meme is funny?

## Dataset Structure

### Files

- `memebench_v1.json`: Annotations for all 1,253 memes
- `metadata.jsonl`: ImageFolder metadata used by the Hugging Face Dataset Viewer. Each row contains `file_name` plus the full VIKR annotation fields.
- `images/`: Directory containing 1,253 meme images (PNG format)

### Annotation Schema

Each entry contains:

| Field | Type | Description |
|-------|------|-------------|
| `id` | int | Unique identifier |
| `image_path` | str | Relative path to meme image |
| `checked_gt` | str | Human-verified ground truth explanation |
| `meta` | object | Metadata for slicing (domain, language, etc.) |
| `moderation` | object | Content safety flags (political, hate_speech) |
| `eval_checklist` | object | VIKR evaluation questions (1-4 per dimension) |
| `visual` | object | Zero-knowledge visual description, OCR, visual entities |
| `identity` | object | Entity identities (name, source) |
| `knowledge` | object | Cultural background facts, cultural code |
| `reasoning` | object | Humor mechanism (trigger, reference, logic, core message) |

### Meta Fields

| Field | Values |
|-------|--------|
| `domain` | ACG, Movies_TV, History, Politics, Daily_Life, Sports, Cross_Domain |
| `meme_type` | Single_Source, Multi_Source, Template |
| `structure` | Single, Comparison, Sequential, Collage |
| `language` | zh, en |

### VIKR Layers

**Visual** (`visual`): Objective visual description with no cultural identification. Includes `description` (scene description), `ocr` (transcribed text), and `entities[]` (visual appearance of each entity).

**Identity** (`identity`): Entity identification with `entities[]` containing `id`, `name` (canonical name), and `source` (origin work/context). Entity IDs are dual-linked with visual entities.

**Knowledge** (`knowledge`): Cultural background with `facts[]` (1-4 essential background facts) and `cultural_code` (specific slang, meme name, or subculture term).

**Reasoning** (`reasoning`): Humor mechanism with `trigger` (surface element), `reference` (deeper reference), `logic` (primary mechanism type), `logic_detail` (compound description), and `core_message` (why the meme works).

## Dataset Statistics

### Domain Distribution

| Domain | Count | Percentage |
|--------|-------|------------|
| ACG | 628 | 50.1% |
| Cross_Domain | 266 | 21.2% |
| Daily_Life | 166 | 13.2% |
| Movies_TV | 125 | 10.0% |
| History | 55 | 4.4% |
| Sports | 7 | 0.6% |
| Politics | 6 | 0.5% |

### Language Distribution

| Language | Count | Percentage |
|----------|-------|------------|
| Chinese (zh) | 768 | 61.3% |
| English (en) | 485 | 38.7% |

### Logic Type Distribution

| Logic Type | Full Name | Count | Percentage |
|------------|-----------|-------|------------|
| IC | Identity Collision | 549 | 43.8% |
| IR | Irony | 364 | 29.0% |
| RV | Rule Violation | 157 | 12.5% |
| SM | Semantic Mismatch | 106 | 8.5% |
| PP | Phonetic Pun | 46 | 3.7% |
| VS | Visual Similarity | 31 | 2.5% |

## Evaluation Protocol

MemeBench uses an LLM-as-Judge evaluation with checklist-based scoring across all four VIKR dimensions. The primary metric is **Complete Pass**: a meme is considered fully understood only when all four dimensions (V=1, I=1, K=1, R=1) are satisfied.

Each dimension is evaluated using 1-4 binary checklist questions. A dimension passes when all its checklist questions are answered correctly.

## Dataset Creation

### Source Data

Memes were collected from Chinese and English internet communities, including forums, social media platforms, and image boards. The collection covers 7 cultural domains with emphasis on ACG (Anime/Comics/Games) and cross-domain content.

### Annotation Process

Annotations were created through a multi-stage pipeline:
1. **LLM-assisted initial annotation**: Structured VIKR fields generated by large language models
2. **Human expert verification**: All entries reviewed and corrected by domain experts
3. **Automated quality audits**: Schema invariants enforced programmatically (entity dual-linkage, cardinality bounds 1-4, layer separation)
4. **Content moderation**: Automated filters with human override for political and hate speech content

### Design Principles

- **Zero-knowledge visual**: Visual layer describes only what is seen, with no cultural identification
- **Entity dual-linkage**: Visual entity IDs and identity entity IDs must correspond one-to-one
- **Knowledge/reasoning separation**: Background knowledge and humor mechanisms are annotated separately
- **Diagnostic-first**: Schema design serves VIKR ablation and evaluation, not just display

## Considerations for Using the Data

### Social Impact

MemeBench enables research on AI understanding of cultural communication through memes. Improving AI comprehension of memes can benefit content moderation, accessibility, and cross-cultural communication.

### Biases

- **Domain bias**: ACG content is overrepresented (50.1%), while Sports (0.6%) and Politics (0.5%) are underrepresented
- **Language bias**: Chinese entries (61.3%) outnumber English entries (38.7%)
- **Cultural bias**: Memes primarily reflect Chinese and English internet culture
- **Temporal bias**: Collection period 2025-2026; older or emerging memes may not be represented

### Limitations

- Evaluation-only benchmark; not intended for model training
- Binary moderation flags provide coarse safety signals
- Cultural knowledge requirements may not generalize across all cultural contexts

### Personal and Sensitive Information

The dataset contains no personally identifiable information (PII). All memes are sourced from publicly available internet content. Some entries reference public figures in satirical or humorous contexts.

## Citation

```bibtex
@inproceedings{anonymous2026memebench,
  title={MemeBench: Diagnosing Cultural-Semantic Understanding in LVLMs through Memes},
  author={Anonymous},
  booktitle={NeurIPS 2026 Evaluations and Datasets Track},
  year={2026}
}
```

## License

This dataset is released under the [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/) license.