merlin-rare

MERLIN-Rare是一个专为评估多语言实体链接中稀有实体性能而设计的数据集。它源自论文《Think Before You Link: Rarity, Reasoning, and Retrieval in Multilingual Entity Linking》，是对MERLIN数据集的扩展。该数据集包含来自五种语言（印地语、印尼语、日语、泰米尔语、越南语）的测试集，总计1105个实体提及和790张唯一图像。每个样本保留了MERLIN的原始字段：文章标题、实体名称、Wikidata ID、英文Wikipedia标题、图像名称（存储为Hugging Face Image类型）和图像文件名。此外，针对15个Wikipedia和Wikidata指标（包括90天页面浏览量、Wikipedia反向链接、文章字节大小、引用数、修订数、独特编辑者、分类数、图片数、外部链接数、Wikidata入链数、Wikidata出链数、语言版本数、陈述数、限定符数、实体年龄天数），每个样本都添加了对应指标的整数值、百分位数以及一个布尔值标识该样本是否属于该指标的底部5%切片。百分位数在每个语言内计算，排除无效的英文Wikipedia标题。数据集的加载可通过Hugging Face Datasets库实现，使用语言代码作为配置。该数据集主要用于多语言实体链接任务中稀有实体的评估和基准测试，配套评估框架可在相关GitHub仓库获取。数据集遵循CC BY-SA 4.0许可证，衍生自MERLIN，Wikipedia文本和图像保留其原始许可证。MERLIN-Rare is a dataset specifically designed for evaluating the performance of rare entities in multilingual entity linking. It originates from the paper Think Before You Link: Rarity, Reasoning, and Retrieval in Multilingual Entity Linking and serves as an extension of the MERLIN dataset. The dataset contains test sets from five languages (Hindi, Indonesian, Japanese, Tamil, Vietnamese), totaling 1,105 entity mentions and 790 unique images. Each sample retains the original fields of MERLIN: Article Title, Entity Name, Wikidata ID, English Wikipedia Title, Image Name (stored as Hugging Face Image type), and Image Filename. Additionally, for 15 Wikipedia and Wikidata metrics (including 90-day page views, Wikipedia backlinks, article byte size, citation count, revision count, unique editors, category count, image count, external link count, Wikidata inbound links, Wikidata outbound links, language version count, statement count, qualifier count, entity age in days), each sample includes integer values, percentiles, and a boolean flag indicating whether the sample falls within the bottom 5% slice of that metric. Percentiles are computed within each language, excluding invalid English Wikipedia titles. The dataset can be loaded via the Hugging Face Datasets library using language codes as configurations. It is primarily used for evaluating and benchmarking rare entities in multilingual entity linking tasks, with an accompanying evaluation framework available in a related GitHub repository. The dataset is licensed under CC BY-SA 4.0, derived from MERLIN, with Wikipedia text and images retaining their original licenses.