EntEval

EntEval是由丰田工业大学芝加哥分校的研究团队开发的一个综合性实体表示评估基准。该数据集包含约8500万条记录，涵盖了多种需要对实体进行非平凡理解的多样化任务，如实体类型识别、实体相似性评估、实体关系预测和实体消歧等。数据集的创建过程中，研究团队利用了维基百科中的自然超链接注释，发展了训练技术以学习更优质的实体表示。EntEval的应用领域广泛，旨在通过标准化评估方法，提升实体表示在语言模型、对话生成、实体链接和故事生成等重要问题中的性能。

EntEval is a comprehensive entity representation evaluation benchmark developed by a research team from Toyota Technological Institute at Chicago. This dataset contains approximately 85 million records, covering a diverse set of tasks that require non-trivial understanding of entities, including entity type recognition, entity similarity evaluation, entity relation prediction, entity disambiguation, and more. During the dataset development process, the research team leveraged natural hyperlink annotations from Wikipedia and developed specialized training techniques to learn higher-quality entity representations. EntEval has broad application scenarios, aiming to enhance the performance of entity representations in key tasks such as language modeling, dialogue generation, entity linking and story generation via standardized evaluation methods.