aiatums/INDOMEME

---
language:
- id
license: cc-by-4.0
task_categories:
- image-classification
- text-classification
tags:
- hate-speech
- meme
- multimodal
- indonesian
- social-media
---

# INDOMEME

INDOMEME is a multimodal dataset of Indonesian memes collected from Facebook, annotated for hate speech detection and content appropriateness classification. Each meme is enriched with OCR-extracted text and LLM-generated captions to support multimodal analysis.

## Dataset Columns

| Column | Description |
|--------|-------------|
| `image` | Meme image |
| `image_path` | Original filename of the meme image |
| `hate_final` | Hatefulness label: `hate` or `not hate` |
| `appropriate_final` | Appropriateness label: `appropriate` or `inappropriate` |
| `topic` | Topical focus in English (e.g., `gender`, `political`, `individual`) |
| `topic_id` | Topical focus in Indonesian |
| `ocr` | Text extracted from the meme image using Qwen2-VL-2B |
| `caption` | Formal image caption generated by Gemini 2.5 Flash describing the visual content and implied message |

### Label Details

**`hate_final`**: A meme is labeled `hate` if it contains insults, harassment, negative stereotypes, or demeaning insinuations toward individuals or groups. Otherwise it is labeled `not hate`.

**`appropriate_final`**: A meme is labeled `inappropriate` if it contains coarse language, sexual references, depictions of violence, or other content unsuitable for public viewing. Otherwise it is labeled `appropriate`. Note that all hateful memes are also labeled as `inappropriate`.

**`topic` / `topic_id`**: Topical focus categories include: Gender, Individual, National Origin/Ethnicity/Race, Political, Religion, Institution/Company, Social Sub-groups, and None/Others. A meme may have multiple topics.

## Citation

If you use this dataset, please cite:

```bibtex
@article{pamungkas2026indomeme,
title = {Decoding hate in memes: multimodal and multitask approaches for low-resource Indonesian social media},
author = {Pamungkas, Endang Wahyu and Wahyuni, Cahyaningtyas Sekar and Amal, Ikhlasul and Purworini, Dian and Rintyarna, Bagus Setya},
journal = {PeerJ Computer Science},
volume = {12},
pages = {e3736},
year = {2026},
doi = {10.7717/peerj-cs.3736}
}
```

---
语言：
- 印尼语（id）
许可协议：CC BY 4.0
任务类别：
- 图像分类
- 文本分类
标签：
- 仇恨言论（hate-speech）
- 表情包（meme）
- 多模态（multimodal）
- 印尼语
- 社交媒体
---

# INDOMEME

INDOMEME是一个源自Facebook的印尼语表情包多模态数据集，经标注用于仇恨言论检测与内容适宜性分类。每个表情包均附带通过光学字符识别（OCR，Optical Character Recognition）提取的文本，以及大语言模型（LLM，Large Language Model）生成的字幕，以支撑多模态分析研究。

## 数据集字段

| 字段 | 描述 |
|--------|-------------|
| `image` | 表情包图像 |
| `image_path` | 表情包图像的原始文件名 |
| `hate_final` | 仇恨性标签：`hate`（仇恨）或`not hate`（非仇恨） |
| `appropriate_final` | 适宜性标签：`appropriate`（适宜）或`inappropriate`（不适宜） |
| `topic` | 英文主题类别（例如`gender`（性别）、`political`（政治）、`individual`（个体）） |
| `topic_id` | 印尼语主题类别 |
| `ocr` | 使用Qwen2-VL-2B模型从表情包图像中提取的文本 |
| `caption` | 由Gemini 2.5 Flash生成的正式图像字幕，用于描述图像视觉内容与隐含信息 |

### 标签说明

**`hate_final`**：若表情包包含针对个人或群体的侮辱、骚扰、负面刻板印象或贬低性暗示，则标注为`hate`（仇恨）；否则标注为`not hate`（非仇恨）。

**`appropriate_final`**：若表情包包含粗俗语言、性暗示、暴力描绘或其他不适宜公开展示的内容，则标注为`inappropriate`（不适宜）；否则标注为`appropriate`（适宜）。需注意，所有仇恨类表情包均会被同时标注为`inappropriate`（不适宜）。

**`topic` / `topic_id`**：主题类别包括：性别（Gender）、个体（Individual）、国籍/族裔/种族（National Origin/Ethnicity/Race）、政治（Political）、宗教（Religion）、机构/企业（Institution/Company）、社会亚群体（Social Sub-groups）以及无/其他（None/Others）。单个表情包可对应多个主题。

## 引用说明

若使用本数据集，请引用以下文献：
bibtex
@article{pamungkas2026indomeme,
title = {Decoding hate in memes: multimodal and multitask approaches for low-resource Indonesian social media},
author = {Pamungkas, Endang Wahyu and Wahyuni, Cahyaningtyas Sekar and Amal, Ikhlasul and Purworini, Dian and Rintyarna, Bagus Setya},
journal = {PeerJ Computer Science},
volume = {12},
pages = {e3736},
year = {2026},
doi = {10.7717/peerj-cs.3736}
}