Increasing Entity Linking upper bound through a more effective Candidate Generation System

Anonymous

Abstract: Entity Linking (EL) aligns entity mentions in text to entries in a knowledge base. It usually comprises of two phases: candidate generation and candidate ranking. While most methods focus on the latter phase, it is candidate generation that sets the upper bound for both time and accuracy of an EL system. We propose a simple approach for improving candidate generation by efficiently embedding mention-entity pairs in dense space through a BERT-based bi-encoder. Specifically, we introduce a new pooling function and incorporate entity type side-information. We achieve a new state-of-the-art 84.28% recall of the gold entity in the Zero-shot EL dataset with just 50 candidates, compared to the previous 82.06% with 64 candidates. We report the results from extensive experimentation using our proposed model on both seen and unseen entity datasets. Our results suggest that our approach could be a useful complement to existing EL methods.