Get the Point! Graph Enhanced Candidate Retrieval for Zero-shot Entity Linking

Abstract

For the retrieval phase of the zero-shot entity linking task, BERT has been widely used to represent the mentions and entities with the sentence embeddings. However, the sentence embeddings obtained by BERT are dominated by the high-frequency words in the pre-training corpus, thus performing poorly especially when the mention/entity is a low-frequency word. To solve this issue, we propose a Graph enhanced Entity Retrieval (GER) framework, fusing word-level embedding with the sentence embedding from BERT. Specifically, we construct a mention/entity centralized graph and design a Hierarchical Graph Neural Network (HGNN) to capture the word-level information. Experimental results on the ZESHEL dataset demonstrate that our proposal achieves a recall@64 of 84.72\%, a 2.66 points improvement compared to previous best results.

Tasks

Reproductions