ner-augmentation
收藏资源简介:
Impresso HIPE-2022 NER 数据集是一个专门用于命名实体识别(NER)任务的数据集,源自 HIPE-2022 共享任务,并为 Impresso NER 模型进行了预处理。数据集的核心内容是从历史报纸文本中提取的命名实体标注数据,涉及德语(de)、英语(en)和法语(fr)三种语言。数据以 token 序列及其对应的 NER 标签形式组织,标注方案采用 IOB2 格式,覆盖了五类粗粒度实体类型:地点(loc)、组织(org)、人物(pers)、产品(prod)和时间(time)。每个数据样本除了 tokens 和 ner_tags 外,还包含元数据字段:日期(date)、来源数据包(bundle,如 hipe2020 或 letemps)和语言(lang),便于按来源和语言进行切片分析。数据集提供了多个配置(或变体),主要围绕 hipe2020 和 letemps 这两个原始数据包及其经过名称替换增强的版本。例如,默认配置 hipe2020-letemps 合并了这两个数据包,包含 23,266 个训练样本、4,832 个开发样本和 5,613 个测试样本。其他配置如 hipe2020-letemps-name-replacement-150 则提供了通过名称替换进行数据增强后规模更大的训练集(58,164 个样本)。数据集适用于历史文本的命名实体识别研究、多语言 NER 模型训练,以及对历史文献进行信息提取等任务。数据遵循 CC BY-NC-SA 4.0 许可协议。
The Impresso HIPE-2022 NER dataset is a specialized dataset for Named Entity Recognition (NER) tasks, derived from the HIPE-2022 shared task and preprocessed for the Impresso NER model. The core content consists of named entity annotations extracted from historical newspaper texts. The data covers three languages: German (de), English (en), and French (fr). It is organized as token sequences with corresponding NER labels, using the IOB2 tagging scheme and covering five coarse-grained entity types: location (loc), organization (org), person (pers), product (prod), and time (time). Each data sample includes tokens and ner_tags, along with metadata fields: date, source bundle (e.g., hipe2020 or letemps), and language (lang), facilitating slice analysis by source and language. The dataset offers multiple configurations or variants, primarily based on the original bundles hipe2020 and letemps and their enhanced versions via name replacement. For example, the default configuration hipe2020-letemps combines these two bundles, containing 23,266 training samples, 4,832 development samples, and 5,613 test samples. Other configurations, such as hipe2020-letemps-name-replacement-150, provide larger training sets (58,164 samples) through data augmentation via name replacement. The dataset is suitable for research on named entity recognition in historical texts, multilingual NER model training, and information extraction from historical documents. The data is licensed under CC BY-NC-SA 4.0.




