Abstract

Named Entity Disambiguation (NED) is the task of mapping textual mentions to entities in a database. A key challenge in NED is generalizing to rarely seen entities, termed tail entities. Traditional NED systems use hand-tuned features to improve tail generalization, but these features make the system challenging to deploy and maintain, especially in multiple locales. In 2018, a subset of the authors built a self-supervised NED system at Apple, which improved performance over its hand-tuned predecessor on a suite of downstream products. Motivated to understand the core reasons for this improvement, we introduce Bootleg, a clean-slate, open-source, self-supervised NED system.1 We first demonstrate that Bootleg matches or exceeds state-of-the-art performance on three NED benchmarks by up to 5.8 F1 points. Importantly, Bootleg improves performance over a BERT-based NED baseline by 41.2 F1 points on tail entities in Wikipedia using a simple transformer-based architecture and a hierarchical regularization scheme. Finally, we observe that embeddings from self-supervised models like Bootleg are increasingly being served to downstream applications, creating an embedding ecosystem. We initiate the study of the data management challenges associated with this ecosystem.