Multiple-source Entity Linking with Incomplete Sources

Qinguo Liu, Shui Liu, Lemao Liu, Bo Xiao

Abstract: This paper introduces a new entity linking task from a well-known online video application in industry, where both entities and mentions are represented by multiple sources but some of them may be missing. To address the issue of incomplete sources, it proposes a novel neural approach to model the linking relationship between a pair of an entity and a mention. To verify the proposed approach to this task, it further creates a large scale dataset including 70k examples. Experiments on this dataset empirically demonstrate that the proposed approach is effective over a baseline and particularly it is robust to the missing sources in some extent.