Improving candidate generation for entity linking

Entity linking is the task of linking names in free text to the referent entities in a knowledge base. Most recently proposed linking systems can be broken down into two steps: candidate generation and candidate ranking. The first step searches candidates from the knowledge base and the second step disambiguates them. Previous works have been focused on the recall of the generation because if the target entity is absent in the candidate set, no ranking method can return the correct result. Most of the recall-driven generation strategies will increase the number of the candidates. However, with large candidate sets, memory/time consuming systems are impractical for online applications. In this paper, we propose a novel candidate generation approach to generate high recall candidate set with small size. Experimental results on two KBP data sets show that the candidate generation recall achieves more than 93%. By leveraging our approach, the candidate number is reduced from hundreds to dozens, the system runtime is saved by 70.3% and 76.6% over the baseline and the highest micro-averaged accuracy in the evaluation is improved by 2.2% and 3.4%.