Wikidata based Person Entity Linking in News Articles
Abdul Lathif Fathima Shanaz, Roshan Ragel · 2021
Entity linking (EL) is a process of extracting entity mentions in documents and linking them to their corresponding actual entities in a Knowledge Base (KB) such as Wikipedia or Wikidata. This task is challenging due to name variations, incompleteness of the KB and high ambiguity of entity mentions. News articles generally contain mentions of entities like persons, organizations, locations, etc. Which are excellent resources for understanding readers’ news interest. However, an entity mention can refer to different real world entities. Furthermore, person entities are generally more ambiguous and critically important to understand the readers’ news interest. This paper aims to design EL system for disambiguating person mentions in news articles. Current EL methods do not focus on improving EL performance of person entity mentions. Wikidata KB is chosen based on accessibility, completeness of the relations and timeliness of the data. The proposed method includes three main steps. In the first step, a new approach is proposed to generate candidate entities based on name dictionary-based technique. Partial string matching technique is adopted when person names are referred part of their full names. In the second step, top N candidates are selected using features from previous studies: entity popularity, textual similarity and the contextual similarity between a mention and the KB entity. In the last step, the best-matched entity is chosen from the top N candidates based on the semantic relatedness between the entities in a news article. The performance of the proposed methods is evaluated over a manually annotated AIDA-CoNLL news dataset. Experimental results show that the proposed approach for candidate entity generation achieves the highest precision and the recall of 98.91% and the top-1 precision of 90.05% at best matched entity selection on AIDA-CoNLL testb dataset.