Building an Entity Linking System without Prior Learning using Wikidata

Abstract

Entity Linking, which links entities in a text to entities in a knowledge graph, has attracted attention as a fundamental technology for information retrieval and question answering systems. In Entity Linking without pre-training, the performance impact of the resources used is significant. This study builds a Japanese Entity Linking system without pre-training using Wikidata as a resource and investigates the performance baseline and challenges in Entity Linking for Wikidata. In addition, it will be verified whether linking performance can be improved by using the Wikidata structure. The results of the verification confirmed that the system is about 60% accurate. In addition, the use of the Wikidata structure was shown to improve linking performance.