chinese-sentiment / README.md
TIX007's picture
更新readme
e58abb3
---
language:
- zh
license: mit
pretty_name: Chinese Emotion Dataset
task_categories:
- text-classification
task_ids:
- sentiment-classification
- multi-class-classification
size_categories:
- 10K<n<100K
tags:
- emotion
- sentiment-analysis
- chinese
- text
- datasets
annotations_creators:
- machine-generated
language_creators:
- machine-generated
source_datasets:
- dair-ai/emotion
dataset_info:
features:
- name: text
dtype: string
- name: label
dtype: int64
- name: label_name
dtype: string
- name: label_cn
dtype: string
splits:
- name: train
num_examples: 16000
- name: validation
num_examples: 2000
- name: test
num_examples: 2000
---
# 中文情感数据集卡片
## 数据集详情
### 数据集简介
中文情感数据集,包含日常生活场景下的中文短文本情感标注。包含 6 种基本情感类别,共 20,000 条样本。
- **策划者:** TIX007
- **语言:** 中文(简体) / zh
- **许可证:** MIT
### 数据来源
- **仓库:** [https://huggingface.co/datasets/TIX007/chinese-sentiment](https://huggingface.co/datasets/TIX007/chinese-sentiment)
## 使用方式
### Direct Use
- 中文情感分类模型的训练与评估
- 文本情感分析(sentiment analysis)研究
- 作为预训练语言模型微调的数据集
- 快速验证中文 NLP 模型的情感识别能力
### Out-of-Scope Use
- 不建议直接用于商业生产环境(文本来源于模板扩增,缺乏真实社交媒体多样性)
- 不适用于细粒度情感分析(如情感强度打分)
- 不适用于对话情感分析或上下文依赖的情感理解
## 数据集结构
### 数据字段
| 字段名 | 类型 | 说明 |
|-------|------|------|
| `text` | string | 中文短文本 |
| `label` | int (0-5) | 情感类别编号 |
| `label_name` | string | 英文情感标签名 |
| `label_cn` | string | 中文情感标签名 |
### 数据划分
| 划分 | 样本数 |
|------|--------|
| train | 16,000 |
| validation | 2,000 |
| test | 2,000 |
| **合计** | **20,000** |
### 标签信息
| 标签 ID | 英文名 | 中文名 |
|--------|--------|--------|
| 0 | sadness | 悲伤 |
| 1 | joy | 快乐 |
| 2 | love | 爱 |
| 3 | anger | 愤怒 |
| 4 | fear | 恐惧 |
| 5 | surprise | 惊讶 |
### 数据文件
```
data/
├── train.csv / train.jsonl / train.json
├── validation.csv / validation.jsonl / validation.json
└── test.csv / test.jsonl / test.json
```
### 示例数据
```json
{"text": "今天的心情好极了,整个人都充满了活力", "label": 1, "label_name": "joy", "label_cn": "快乐"}
{"text": "眼泪悄悄滑落,却没有人在身旁安慰", "label": 0, "label_name": "sadness", "label_cn": "悲伤"}
{"text": "每次看到你,我的心就变得很柔软", "label": 2, "label_name": "love", "label_cn": "爱"}
{"text": "简直无法理解这种行为,太令人气愤了", "label": 3, "label_name": "anger", "label_cn": "愤怒"}
{"text": "一想到那种情景,就浑身发抖", "label": 4, "label_name": "fear", "label_cn": "恐惧"}
{"text": "这个消息真的把我惊呆了,半天没回过神来", "label": 5, "label_name": "surprise", "label_cn": "惊讶"}
```
### 数据加载方式
```python
from datasets import load_dataset
dataset = load_dataset(
"csv",
data_files={
"train": "data/train.csv",
"validation": "data/validation.csv",
"test": "data/test.csv",
}
)
print(dataset)
```
## 数据集创建
### 策划说明
构建一个覆盖日常场景的中文情感分类数据集,用于中文 NLP 模型的情感识别任务训练与评估。
### 源数据
#### 数据收集与处理
- 为每种情感类别人工编写 100+ 条代表性模板句子,覆盖日常生活、社交、情感表达等场景
- 通过添加时间状语前缀和语气后缀进行多样化扩增
- 使用同义词替换、句式变换等方法增加文本多样性
- 使用固定随机种子(seed=42)保证可复现性
#### 源数据生产者
由人工编写的模板句子,经程序化扩增生成。
### 标注信息
#### 标注过程
标签由模板句子所属的情感类别自动确定,无需额外人工标注。
#### 标注者
机器生成(machine-generated),基于预定义的情感类别模板。
#### 个人与敏感信息
数据集不包含任何个人信息或敏感内容。所有文本均为通用场景描述。
## 偏差、风险与局限性
- 文本来源于模板扩增,风格较为统一,缺乏真实社交媒体的多样性
- 类别分布不完全均衡,训练时建议使用加权损失函数
- 情感标签为粗粒度分类,无法表达混合情感或情感强度
### 建议
- 适合作为预训练微调或快速验证使用
- 如需更高真实度,建议与真实社交媒体数据混合训练
- 生产环境使用前建议在真实数据上进行额外评估
## 引用
```bibtex
@misc{chinese-emotion-dataset-2026,
title = {Chinese Emotion Dataset},
author = {TIX007},
howpublished = {\url{https://huggingface.co/datasets/TIX007/chinese-sentiment}},
note = {中文日常短文本情感分类数据集,MIT 许可证},
year = {2026},
}
```
## 数据集卡作者
TIX007
## 数据集卡联系人
[TIX007](https://huggingface.co/TIX007)