Bridging Semantic and Geographic Space : Embedding-based Toponym Resolution in Historical Guidebooks

Sammanfattning

This thesis investigates the effectiveness of using contextual vector embeddings to disambiguate toponyms found in historical guidebooks. This is a crucial task in the study of tourism patterns throughout history, and to gain a deeper understanding of toponym resolution methods on this specific genre of text. The research here attempts to leverage the connection between semantic similarity and geographic proximity, to investigate the extent to which contextual embeddings can be used to disambiguate toponyms in German historical guidebooks. During the study, candidate vector embeddings were created from geographical databases, which were then compared with semantic embeddings from the context of each marked toponym within the guidebook. Results show high similarity between embeddings created from geographic information and those created from guidebook context, that can help to narrow down potential candidates for disambiguation. However, it also presents unique challenges that must be overcome by other methods in order to produce consistent results.