--- license: cc-by-sa-4.0 task_categories: - summarization language: - zh tags: - chemistry - biology - finance - legal - music - art - code - climate - medical - synthetic pretty_name: CAMS size_categories: - 1M 📷 CAMS: 一个大规模、多方面、基于属性的中文摘要数据集

💻 Github Repo
简体中文 | English

--- ## 简介 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。 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。 ### 主要特点 - **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。 - **多层次摘要**: 每篇文章都提供三个层级结构的摘要: - **长摘要 (Long Summary)**: 详细、全面地覆盖原文的关键信息。 - **中摘要 (Medium Summary)**: 简洁地概括文章的核心要点。 - **短摘要 (Short Summary)**: 一句话总结文章的中心思想。 - **丰富的属性标注**: 每篇文章都经过了多维度属性标注,包括: - **关键词 (Keywords)** - **陈述类型 (Statement Type)**: 事实陈述 vs. 观点表达 - **情感倾向 (Sentiment)**: 正面、较为正面、中性、较为负面、负面 - **正式度 (Formality)**: 正式文体 vs. 口语化文体 - **时态 (Tense)**: 过去、现在、将来 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。 ## 📊 数据统计 ### 基本信息 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。
数据集 规模 平均文章长度 平均摘要长度 平均关键词数目
英文
NYT 655K 552.1 42.8 -
CNNDM 312K 791.7 55.2 -
Newsroom 1.0M 765.6 30.2 -
中文
LCSTS 2.4M 103.7 17.9 -
CLTS 185K 1363.7 58.1 -
CNewSum 396K 730.4 35.1 -
CSL 396K 206.0 19.0 5.1
CAMS 1.0M 1571.4 60.0 (S)
185.7 (M)
428.1 (L)
14.3
### 主题分布 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: ```json { 'other_manufacturing': 'Other Manufacturing', 'automobile': 'Automobile', 'biomedicine': 'Biomedicine', 'computer_communication': 'Computer Communication', 'subject_education_education': 'Education', 'finance_economics': 'Finance & Economics', 'transportation': 'Transportation', 'literature_emotion': 'Literature & Emotion', 'water_resources_ocean': 'Water Resources', 'aerospace': 'Aerospace', 'technology_scientific_research': 'Scientific Research', 'electric_power_energy': 'Energy & Power', 'mining': 'Mining', 'petrochemical': 'Petrochemical', 'law_judiciary': 'Law & Judiciary', 'accommodation_catering_hotel': 'Hospitality', 'film_entertainment': 'Film & Entertainment', 'agriculture_forestry_animal_husbandry_fishery': 'Agriculture & Fishery', 'current_affairs_government_administration': 'Government & Public Affairs', 'news_media': 'News & Media', 'artificial_intelligence_machine_learning': 'AI & Machine Learning', 'computer_programming_code': 'Software Development', 'sports': 'Sports', 'fire_safety_food_safety': 'Food & Fire Safety', 'mathematics_statistics': 'Math & Statistics', 'medicine_health_psychology_traditional_chinese_medicine': 'Medicine & Health', 'game': 'Gaming', 'other_information_services_information_security': 'Information Security', 'real_estate_construction': 'Real Estate & Construction', 'tourism_geography': 'Tourism & Geography' } ``` 具体样本主题分布如下: ![Topic Distribution](static/topic_distribution.jpg) 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: ![Topic Visualization](static/topic_visualization.jpg) ### 属性标注 我们展示了四种额外标注信息的分布: ![Attributes](static/attributes.jpg) ## 📂 数据格式 数据集中的每个样本都以 JSON 格式存储,包含以下字段: ```json { "id": "每条数据的唯一标识", "text": "文章的原始内容", "topic": "文章的主题", "short_summary": "一句话短摘要", "medium_summary": "中等长度的摘要", "long_summary": "详细的长摘要", "keywords": ["关键词1", "关键词2", "关键词3", "..."], "statement_type": "是否为事实陈述", "sentiment": "文章写作情感/作者立场情感", "formality": "文章写作风格正式度", "tense": "文章写作时态", } ``` ## 🛠️ 数据集构建 CAMS 的构建过程主要分为三个阶段: 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。 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。 Stepwise Generation Pipeline 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。 ## 🚀 使用示例 您可以使用 🤗 `datasets` 库轻松加载 CAMS 数据集。 ```python from datasets import load_dataset # 加载 CAMS 数据集 dataset = load_dataset("Mxode/CAMS") # 查看数据集结构 print(dataset) # 读取第一个样本 sample = dataset["train"][0] print("文章:", sample["text"][:200]) print("短摘要:", sample["short_summary"]) print("关键词:", sample["keywords"]) ``` ## 📜 引用 如果您在研究中使用了 CAMS 数据集,请引用我们的工作: ```bibtex @misc{zhang2025CAMS, title={CAMS: A Large-Scale Chinese Attribute-based Multi-faceted Summarization Dataset}, url={https://huggingface.co/datasets/Mxode/CAMS}, author={Xiantao Zhang}, month={August}, year={2025} } ``` ## 📄 许可 本数据集采用 **CC BY-SA 4.0** 许可。