Elgold partial: News

The dataset contains 37 English texts scrapped from news websites. In each text, the named entities are marked. Each name entity is linked to the corresponding Wikipedia if possible. All entities were manually verified by at least three people, which makes the dataset a high-quality gold standard for the evaluation of named entity recognition and linking algorithms.
Each marked entity in the dataset is assigned to one of the following classes:EVENT - Named hurricanes, battles, wars, sports events, etc.FAC - Buildings, airports, highways, bridges, etc.GPE - Countries, cities, statesLANGUAGE - Any named languageLAW - Named documents made into laws.LOC - Non-GPE locations, mountain ranges, bodies of waterNORP - Nationalities or religious or political groupsORG - Companies, agencies, institutions, etc.PERSON - People, including fictionalPRODUCT - Objects, vehicles, foods, etc. (not services)WORK_OF_ART - Titles of books, songs, etc.DISEASE - Names of diseasesSUBSTANCE - Natural substancesSPECIE - Species names of animals, plants, viruses, etc.
The marked entities are embedded directly in the textual files using the following syntax:
{{mention content|entity class|Wikipedia target}}
The "mention content" is a fragment of the textual file that was marked, "entity class" is the named entity class, and "Wikipedia target" is the normalized name of the English Wikipedia page describing the entity. If the entity cannot be linked sensibly to any article the target is empty but the second pipe (|) is preserved. 
There is a guarantee that the double braces in the texts exist only as marked entity syntax. It allows to process the files using simple regular expression: {{[^{}]*}} The distribution of datast NER classes, split into separate categories. The first number shows the quantity of linked entities, the second all marked mentions.