Named Entity Extraction of Wikidata Items

Abstract

Wikidata is a free collaborative knowledge base with more than 56 million data items. It contains structured knowledge of named entities such as people, locations, events, etc. If we can link the entities in text documents to corresponding Wikidata items, we can semantically understand the text. To accomplish this task, we must first perform the classification of Wikidata entities into predefined entity types. In this paper, we propose an enhanced approach for assigning entity types to Wikidata entities. This method outputs a named entity dataset containing classification of Wikidata entities to the three main entity classes (Person, Location and Organization), which are widely used in Natural Language Processing. This provides classification of over 12 million Wikidata entities to the three main entity classes. The results are compared to previous work in the field. We observe large improvement of 5.45% in overall precision and 2.09% in F1-score. Our method categorizes Wikidata entities into the three main classes with precision of 95.15% and F1-score of 90.09%. Based on this, we extracted and classified 12,304,719 Wikidata entities, of which 6,635,444 are locations (53.93%), 4,445,785 persons (36.13%), and 1,223,490 organizations (9.94%).