A pipeline Japanese entity linking system with embedding features

Entity linking (EL) is the task of connecting mentions in texts to entities in a large-scale knowledge base such asWikipedia. In this paper, we present a pipeline system for Japanese EL which consists of two standard components, namely candidate generation and candidate ranking. We investigate several techniques for each component, using a recently developed Japanese EL corpus. For candidate generation, we find that a concept dictionary using anchor texts of Wikipedia is more effective than methods based on surface similarity. For candidate ranking, we verify that a set of features used in English EL is effective in Japanese EL as well. In addition, by using a corpus that links Japanese mentions to Japanese Wikipedia entries, we are able to get rich context information from Japanese Wikipedia articles and benefit mention disambiguation. It was not directly possible with previous EL corpora, which associate mentions to English Wikipedia entities. We take this advantage by exploring several embedding models that encode context information of Wikipedia entities, and show that they improve candidate ranking. As a whole, our system achieves 82.27% accuracy, significantly outperforming previous work.