Zero-Shot Entity Typing in Knowledge Graphs

Knowledge graphs are often highly incomplete due to their large sizes and one major task for knowledge graph completion is entity typing, that is to predict missing types of entities or vice versa. It is especially challenging to perform entity typing when the type is new, i.e., unseen during training, which is known as the zero-shot entity typing problem. Existing entity typing models cannot handle the zero-shot case as it requires the models to be retrained to embed the unseen types, and other zero-shot knowledge graph completion approaches cannot be applied to the entity typing task either. In this paper, we propose a novel zero-shot entity typing approach based on a generation architecture, and introduce a novel feature distribution and semantic encoding method that combines both ontological and textual knowledge. We also construct the first zero-shot entity typing datasets based on commonly used benchmarks. Our experiment evaluation shows the effectiveness of our approach.