Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,283 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: cc-by-sa-4.0
|
| 3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: cc-by-sa-4.0
|
| 3 |
+
task_categories:
|
| 4 |
+
- summarization
|
| 5 |
+
language:
|
| 6 |
+
- zh
|
| 7 |
+
tags:
|
| 8 |
+
- chemistry
|
| 9 |
+
- biology
|
| 10 |
+
- finance
|
| 11 |
+
- legal
|
| 12 |
+
- music
|
| 13 |
+
- art
|
| 14 |
+
- code
|
| 15 |
+
- climate
|
| 16 |
+
- medical
|
| 17 |
+
- synthetic
|
| 18 |
+
pretty_name: CAMS
|
| 19 |
+
size_categories:
|
| 20 |
+
- 1M<n<10M
|
| 21 |
+
---
|
| 22 |
+
|
| 23 |
+
<h1 align="center">
|
| 24 |
+
📷 CAMS: 一个大规模、多方面、基于属性的中文摘要数据集
|
| 25 |
+
</h1>
|
| 26 |
+
|
| 27 |
+
<p align="center">
|
| 28 |
+
<a href="https://github.com/Mxoder/Maxs-Awesome-Datasets" target="_blank">💻 Github Repo</a> <br>
|
| 29 |
+
简体中文 | <a href="https://huggingface.co/datasets/Mxode/CAMS/blob/main/README-en.md" target="_blank">English</a> <br>
|
| 30 |
+
</p>
|
| 31 |
+
|
| 32 |
+
---
|
| 33 |
+
|
| 34 |
+
## 简介
|
| 35 |
+
|
| 36 |
+
CAMS (**C**hinese **A**ttribute-based **M**ulti-faceted **S**ummarization) 是一个为推进长文本摘要研究而设计的大规模中文摘要数据集。随着大型语言模型(LLMs)的快速发展,高质量、大规模的训练数据变得至关重要,特别是在非英语语种中。CAMS 旨在填补中文长文本摘要领域的空白。
|
| 37 |
+
|
| 38 |
+
数据集包含 **100万篇** 高质量的中文长文章,每篇文章都配有三个不同粒度的摘要和一组丰富的属性标签。
|
| 39 |
+
|
| 40 |
+
### 主要特点
|
| 41 |
+
|
| 42 |
+
- **专注长文本**: 数据集中的文章平均长度超过 1500 个字符,为长文本摘要模型提供了富有挑战性的训练和评估平台。
|
| 43 |
+
- **多层次摘要**: 每篇文章都提供三个层级结构的摘要:
|
| 44 |
+
- **长摘要 (Long Summary)**: 详细、全面地覆盖原文的关键信息。
|
| 45 |
+
- **中摘要 (Medium Summary)**: 简洁地概括文章的核心要点。
|
| 46 |
+
- **短摘要 (Short Summary)**: 一句话总结文章的中心思想。
|
| 47 |
+
- **丰富的属性标注**: 每篇文章都经过了多维度属性标注,包括:
|
| 48 |
+
- **关键词 (Keywords)**
|
| 49 |
+
- **陈述类型 (Statement Type)**: 事实陈述 vs. 观点表达
|
| 50 |
+
- **情感倾向 (Sentiment)**: 正面、较为正面、中性、较为负面、负面
|
| 51 |
+
- **正式度 (Formality)**: 正式文体 vs. 口语化文体
|
| 52 |
+
- **时态 (Tense)**: 过去、现在、将来
|
| 53 |
+
|
| 54 |
+
我们希望 CAMS 数据集能够推动可控摘要、属性感知生成和长文本理解等领域的研究与创新。
|
| 55 |
+
|
| 56 |
+
|
| 57 |
+
|
| 58 |
+
|
| 59 |
+
|
| 60 |
+
## 📊 数据统计
|
| 61 |
+
|
| 62 |
+
### 基本信息
|
| 63 |
+
|
| 64 |
+
为了更好地展示 CAMS 的基本信息,我们将其与其他主流的中英文摘要数据集进行了比较。
|
| 65 |
+
|
| 66 |
+
<table>
|
| 67 |
+
<thead>
|
| 68 |
+
<tr>
|
| 69 |
+
<th style="text-align: left;"><strong>数据集</strong></th>
|
| 70 |
+
<th style="text-align: left;"><strong>规模</strong></th>
|
| 71 |
+
<th style="text-align: left;"><strong>平均文章长度</strong></th>
|
| 72 |
+
<th style="text-align: left;"><strong>平均摘要长度</strong></th>
|
| 73 |
+
<th style="text-align: left;"><strong>平均关键词数目</strong></th>
|
| 74 |
+
</tr>
|
| 75 |
+
</thead>
|
| 76 |
+
<tbody>
|
| 77 |
+
<tr>
|
| 78 |
+
<td colspan="5" style="text-align: center;"><strong>英文</strong></td>
|
| 79 |
+
</tr>
|
| 80 |
+
<tr>
|
| 81 |
+
<td style="text-align: left;">NYT</td>
|
| 82 |
+
<td style="text-align: left;">655K</td>
|
| 83 |
+
<td style="text-align: left;">552.1</td>
|
| 84 |
+
<td style="text-align: left;">42.8</td>
|
| 85 |
+
<td style="text-align: left;">-</td>
|
| 86 |
+
</tr>
|
| 87 |
+
<tr>
|
| 88 |
+
<td style="text-align: left;">CNNDM</td>
|
| 89 |
+
<td style="text-align: left;">312K</td>
|
| 90 |
+
<td style="text-align: left;">791.7</td>
|
| 91 |
+
<td style="text-align: left;">55.2</td>
|
| 92 |
+
<td style="text-align: left;">-</td>
|
| 93 |
+
</tr>
|
| 94 |
+
<tr>
|
| 95 |
+
<td style="text-align: left;">Newsroom</td>
|
| 96 |
+
<td style="text-align: left;">1.0M</td>
|
| 97 |
+
<td style="text-align: left;">765.6</td>
|
| 98 |
+
<td style="text-align: left;">30.2</td>
|
| 99 |
+
<td style="text-align: left;">-</td>
|
| 100 |
+
</tr>
|
| 101 |
+
<tr>
|
| 102 |
+
<td colspan="5" style="text-align: center;"><strong>中文</strong></td>
|
| 103 |
+
</tr>
|
| 104 |
+
<tr>
|
| 105 |
+
<td style="text-align: left;">LCSTS</td>
|
| 106 |
+
<td style="text-align: left;">2.4M</td>
|
| 107 |
+
<td style="text-align: left;">103.7</td>
|
| 108 |
+
<td style="text-align: left;">17.9</td>
|
| 109 |
+
<td style="text-align: left;">-</td>
|
| 110 |
+
</tr>
|
| 111 |
+
<tr>
|
| 112 |
+
<td style="text-align: left;">CLTS</td>
|
| 113 |
+
<td style="text-align: left;">185K</td>
|
| 114 |
+
<td style="text-align: left;">1363.7</td>
|
| 115 |
+
<td style="text-align: left;">58.1</td>
|
| 116 |
+
<td style="text-align: left;">-</td>
|
| 117 |
+
</tr>
|
| 118 |
+
<tr>
|
| 119 |
+
<td style="text-align: left;">CNewSum</td>
|
| 120 |
+
<td style="text-align: left;">396K</td>
|
| 121 |
+
<td style="text-align: left;">730.4</td>
|
| 122 |
+
<td style="text-align: left;">35.1</td>
|
| 123 |
+
<td style="text-align: left;">-</td>
|
| 124 |
+
</tr>
|
| 125 |
+
<tr>
|
| 126 |
+
<td style="text-align: left;">CSL</td>
|
| 127 |
+
<td style="text-align: left;">396K</td>
|
| 128 |
+
<td style="text-align: left;">206.0</td>
|
| 129 |
+
<td style="text-align: left;">19.0</td>
|
| 130 |
+
<td style="text-align: left;">5.1</td>
|
| 131 |
+
</tr>
|
| 132 |
+
<tr>
|
| 133 |
+
<td style="text-align: left;"><strong>CAMS</strong></td>
|
| 134 |
+
<td style="text-align: left;"><strong>1.0M</strong></td>
|
| 135 |
+
<td style="text-align: left;"><strong>1571.4</strong></td>
|
| 136 |
+
<td style="text-align: left;"><strong>60.0 (S)</strong> <br><strong>185.7 (M)</strong> <br><strong>428.1 (L)</strong></td>
|
| 137 |
+
<td style="text-align: left;"><strong>14.3</strong></td>
|
| 138 |
+
</tr>
|
| 139 |
+
</tbody>
|
| 140 |
+
</table>
|
| 141 |
+
|
| 142 |
+
### 主题分布
|
| 143 |
+
|
| 144 |
+
CAMS 总共有 30 类不同的主题,如下所示,其中 key 为数据集内字段,value 为主题内容:
|
| 145 |
+
|
| 146 |
+
```json
|
| 147 |
+
{
|
| 148 |
+
'other_manufacturing': 'Other Manufacturing',
|
| 149 |
+
'automobile': 'Automobile',
|
| 150 |
+
'biomedicine': 'Biomedicine',
|
| 151 |
+
'computer_communication': 'Computer Communication',
|
| 152 |
+
'subject_education_education': 'Education',
|
| 153 |
+
'finance_economics': 'Finance & Economics',
|
| 154 |
+
'transportation': 'Transportation',
|
| 155 |
+
'literature_emotion': 'Literature & Emotion',
|
| 156 |
+
'water_resources_ocean': 'Water Resources',
|
| 157 |
+
'aerospace': 'Aerospace',
|
| 158 |
+
'technology_scientific_research': 'Scientific Research',
|
| 159 |
+
'electric_power_energy': 'Energy & Power',
|
| 160 |
+
'mining': 'Mining',
|
| 161 |
+
'petrochemical': 'Petrochemical',
|
| 162 |
+
'law_judiciary': 'Law & Judiciary',
|
| 163 |
+
'accommodation_catering_hotel': 'Hospitality',
|
| 164 |
+
'film_entertainment': 'Film & Entertainment',
|
| 165 |
+
'agriculture_forestry_animal_husbandry_fishery': 'Agriculture & Fishery',
|
| 166 |
+
'current_affairs_government_administration': 'Government & Public Affairs',
|
| 167 |
+
'news_media': 'News & Media',
|
| 168 |
+
'artificial_intelligence_machine_learning': 'AI & Machine Learning',
|
| 169 |
+
'computer_programming_code': 'Software Development',
|
| 170 |
+
'sports': 'Sports',
|
| 171 |
+
'fire_safety_food_safety': 'Food & Fire Safety',
|
| 172 |
+
'mathematics_statistics': 'Math & Statistics',
|
| 173 |
+
'medicine_health_psychology_traditional_chinese_medicine': 'Medicine & Health',
|
| 174 |
+
'game': 'Gaming',
|
| 175 |
+
'other_information_services_information_security': 'Information Security',
|
| 176 |
+
'real_estate_construction': 'Real Estate & Construction',
|
| 177 |
+
'tourism_geography': 'Tourism & Geography'
|
| 178 |
+
}
|
| 179 |
+
```
|
| 180 |
+
|
| 181 |
+
具体样本主题分布如下:
|
| 182 |
+
|
| 183 |
+

|
| 184 |
+
|
| 185 |
+
我们也同样抽取了一部分样本,获取文本嵌入后利用 UMAP 降维可视化了主题分布:
|
| 186 |
+
|
| 187 |
+

|
| 188 |
+
|
| 189 |
+
### 属性标注
|
| 190 |
+
|
| 191 |
+
我们展示了四种额外标注信息的分布:
|
| 192 |
+
|
| 193 |
+

|
| 194 |
+
|
| 195 |
+
|
| 196 |
+
|
| 197 |
+
|
| 198 |
+
|
| 199 |
+
## 📂 数据格式
|
| 200 |
+
|
| 201 |
+
数据集中的每个样本都以 JSON 格式存储,包含以下字段:
|
| 202 |
+
|
| 203 |
+
```json
|
| 204 |
+
{
|
| 205 |
+
"id": "每条数据的唯一标识",
|
| 206 |
+
"text": "文章的原始内容",
|
| 207 |
+
"topic": "文章的主题",
|
| 208 |
+
"short_summary": "一句话短摘要",
|
| 209 |
+
"medium_summary": "中等长度的摘要",
|
| 210 |
+
"long_summary": "详细的长摘要",
|
| 211 |
+
"keywords": ["关键词1", "关键词2", "关键词3", "..."],
|
| 212 |
+
"statement_type": "是否为事实陈述",
|
| 213 |
+
"sentiment": "文章写作情感/作者立场情感",
|
| 214 |
+
"formality": "文章写作风格正式度",
|
| 215 |
+
"tense": "文章写作时态",
|
| 216 |
+
}
|
| 217 |
+
```
|
| 218 |
+
|
| 219 |
+
|
| 220 |
+
|
| 221 |
+
|
| 222 |
+
|
| 223 |
+
## 🛠️ 数据集构建
|
| 224 |
+
|
| 225 |
+
CAMS 的构建过程主要分为三个阶段:
|
| 226 |
+
|
| 227 |
+
1. **数据源与预处理**: 我们从一个大规模、高质量的行业语料库 [IndustryCorpus2](https://huggingface.co/datasets/BAAI/IndustryCorpus2) 中筛选出约 1000 万篇文章作为初始候选集,并经过严格的质量过滤、启发式规则过滤和主题平衡性重采样,最终得到 100 万篇高质量、主题多样的文章。
|
| 228 |
+
|
| 229 |
+
2. **多层次摘要生成**: 我们提出了一种 **逐步生成(Stepwise Generation)** 流程。该流程首先从原文生成详细的长摘要,然后将原文和长摘要作为上下文来生成中摘要,最后再结合原文、长摘要和中摘要来生成最精炼的短摘要。这种方法确保了不同层次摘要之间的一致性和连贯性。
|
| 230 |
+
|
| 231 |
+
<img src="static/pipeline.jpg" alt="Stepwise Generation Pipeline" style="zoom:50%;" />
|
| 232 |
+
|
| 233 |
+
3. **多方面属性标注**: 我们对每篇文章进行关键词提取和多个语言学、文体学维度的属性标注,并通过多轮生成和投票机制确保标注的准确性。
|
| 234 |
+
|
| 235 |
+
|
| 236 |
+
|
| 237 |
+
|
| 238 |
+
|
| 239 |
+
## 🚀 使用示例
|
| 240 |
+
|
| 241 |
+
您可以使用 🤗 `datasets` 库轻松加载 CAMS 数据集。
|
| 242 |
+
|
| 243 |
+
```python
|
| 244 |
+
from datasets import load_dataset
|
| 245 |
+
|
| 246 |
+
# 加载 CAMS 数据集
|
| 247 |
+
dataset = load_dataset("Mxode/CAMS")
|
| 248 |
+
|
| 249 |
+
# 查看数据集结构
|
| 250 |
+
print(dataset)
|
| 251 |
+
|
| 252 |
+
# 读取第一个样本
|
| 253 |
+
sample = dataset["train"][0]
|
| 254 |
+
print("文章:", sample["text"][:200])
|
| 255 |
+
print("短摘要:", sample["short_summary"])
|
| 256 |
+
print("关键词:", sample["keywords"])
|
| 257 |
+
```
|
| 258 |
+
|
| 259 |
+
|
| 260 |
+
|
| 261 |
+
|
| 262 |
+
|
| 263 |
+
## 📜 引用
|
| 264 |
+
|
| 265 |
+
如果您在研究中使用了 CAMS 数据集,请引用我们的工作:
|
| 266 |
+
|
| 267 |
+
```bibtex
|
| 268 |
+
@misc{BiST,
|
| 269 |
+
title={CAMS: A Large-Scale Chinese Attribute-based Multi-faceted Summarization Dataset},
|
| 270 |
+
url={https://huggingface.co/datasets/Mxode/CAMS},
|
| 271 |
+
author={Xiantao Zhang},
|
| 272 |
+
month={August},
|
| 273 |
+
year={2025}
|
| 274 |
+
}
|
| 275 |
+
```
|
| 276 |
+
|
| 277 |
+
|
| 278 |
+
|
| 279 |
+
|
| 280 |
+
|
| 281 |
+
## 📄 许可
|
| 282 |
+
|
| 283 |
+
本数据集采用 **CC BY-SA 4.0** 许可。
|