E3L: Experience Enhanced Entity Linking for Question Answering over Knowledge Graphs

Zhirong Hou,Meiling Wang,Min Li,Ying Li

DOI: https://doi.org/10.1007/978-3-031-25198-6_17

2023-01-01

Abstract:Entity linking is a pivotal factor for building robust Question Answering systems over Knowledge Graphs (KGQA), and representation of entities occupies an important position for tackling entity linking for questions. To alleviate the deficiency of entity descriptions that contextual knowledge is insufficient for entity representations, we introduce entity experiences as a new text-style contextual knowledge source to enrich entity representations, and propose an Experience Enhanced Entity Linking framework called as E 3 L . For the modeling of entity experiences, we derive embeddings using the entity mentions occurring in history questions, and design an attention-based retriever to capture key information relative to user questions. Then the distilled entity experiences are integrated with entity descriptions to enhance entity representations, and question representations are refined with a multi-level attention mechanism. Finally, entity linking is improved with the entity representations and the refined representations of questions. Experimental results on end-to-end benchmark datasets demonstrate that our approach achieves state-of-the-art F1-score, and provides an effective way to improve test performance for universal models using entity experiences without fine-tuning.