Statistical and Neural Methods for Cross-lingual Entity Label Mapping in Knowledge Graphs

ParaNames: A Massively Multilingual Entity Name Corpus

“…One line of research has focused on evaluating the quality of Wikidata-derived labels. For instance, Amaral et al (2022) showed that the "also-knownas" metadata may contain more useful labels than the main labels that Wikidata provides, and that sentence embeddings can be used to pick the best label for each pair in a way that improves the degree to which the names are parallel. However, their experimental results focus on a manually annotated dataset of 10 languages, which allowed them to do much more manual annotation and quality control than in our work.…”