Entity Linking for KGQA Using AMR Graphs

Abstract

Entity linking is an essential part of analytical systems for question answering on knowledge graphs (KGQA). The mentioned entity has to be spotted in the text and linked to the correct resource in the knowledge graph (KG). With this paper, we present our approach on entity linking using the abstract meaning representation (AMR) of the question to spot the surface forms of entities. We re-trained AMR models with automatically generated training data. Based on these models, we extract surface forms and map them to an entity dictionary of the desired KG. For the disambiguation process, we evaluated different options and configurations on QALD-9 and LC-QuaD 2.0. The results of the best performing configurations outperform existing entity linking approaches.