I, RE:Claudius 256: Towards Linking Classical Latin Person Mentions to a Domain-specific Knowledge Base

Abstract

This paper considers Named Entity Linking for person mentions from classical Latin texts to a domain-specific, German language knowledge base, namely Paulys RealencyclopΣdie. Following a methodology similar to (anonymous_reference), we train a transformer-based, retrieval and ranking model (BLINK) first on a general, Wikipedia-derived dataset and subsequently on a more specific dataset, gathered from various sources, linking to our target knowledge base. Results show that while BLINK performs well on mention-entity pairs linked to entities seen during training, it performs significantly worse on mention-entity pairs linking to unseen entities. We provide a detailed error analysis, propose possible exploitation strategies for a human-in-the-loop approach, and identify directions for future improvement.

Resources

Details