TIB-SID (TIB Subject Indexing Dataset)
收藏资源简介:
TIB-SID是由TIB莱布尼茨中心主导构建的双语(英语/德语)多领域图书馆目录数据集,包含13.6万条标注德国规范文档(GND)主题词的编目记录,涵盖文章、书籍、会议论文等五大文献类型。数据集采用JSON-LD格式存储,包含标题、摘要等元数据,并预置训练集/验证集/测试集划分。其特色在于将大规模文本分类任务与权威知识组织系统结合,通过GND的20.7万个主题词及其层级关系,支持面向图书馆实际需求的语义索引研究,适用于多语言主题映射、检索增强分类等数字图书馆AI应用场景。
TIB-SID is a bilingual (English/German) multi-domain library catalog dataset developed and led by the TIB Leibniz Centre. It contains 136,000 catalog records annotated with subject terms from the German National Authority File (GND), covering five types of literature including journal articles, books, conference papers and others. The dataset is stored in JSON-LD format, includes metadata such as titles and abstracts, and has pre-defined training, validation and test set splits. A distinctive feature of TIB-SID is its integration of large-scale text classification tasks with authoritative knowledge organization systems: by leveraging the 207,000 GND subject terms and their hierarchical relationships, it supports semantic indexing research tailored to the actual operational needs of libraries, and is applicable to AI application scenarios in digital libraries such as multilingual subject mapping and retrieval-augmented classification.
TIB-SID 数据集概述
数据集名称
TIB Subject Indexing Dataset (TIB-SID)
核心定位
TIB-SID 是一个用于极端多标签文本分类 (XMTC) 的双语基准数据集,基于真实的图书馆记录。它专为领域分类和基于GND的主题标引而设计。该数据集结合了大规模、结构化、受控的标签空间,以及长尾稀疏性、跨语言差异和现实世界的领域不平衡,使其比标准的文本分类基准更接近实际的图书馆编目操作。
关键数据指标
- 记录数量:136,569 条图书馆记录
- 数据格式:JSON-LD
- 基准划分:包含预定义的训练集、开发集和测试集划分
- 语言:英语和德语
- 领域数量:28个(领域列表文件:https://github.com/sciknoworg/tib-sid/blob/main/28_domains_list.csv)
- 记录类型:文章、书籍、会议文献、报告、学位论文
数据获取
数据集下载地址:https://github.com/sciknoworg/tib-sid/tree/main/library-records-dataset/data
相关背景
TIB-SID 通过 LLMs4Subjects 共享任务(于2025年组织)被引入。全球参与团队在该数据集上开发并评估了超过12个基于LLM的系统。相关共享任务网站提供了更多背景、任务详情和排行榜结果。
- LLMs4Subjects @ SemEval: https://sites.google.com/view/llms4subjects
- LLMs4Subjects @ GermEval: https://sites.google.com/view/llms4subjects-germeval/
引用信息
引用信息即将发布。
许可协议
本作品采用 知识共享署名-相同方式共享 4.0 国际许可协议 进行许可。
- 1An Extreme Multi-label Text Classification (XMTC) Library Dataset: What if we took "Use of Practical AI in Digital Libraries" seriously?TIB莱布尼茨科学与技术信息中心; 德国国家图书馆; 乌迪内大学; 芬兰国家图书馆 · 2026年



