Zero-Shot Entity Representation Learning for Temporal Knowledge Graph

Temporal Knowledge Graphs (TKG) are multi-relational graphs where time is an important dimension. The research interest in TKG is increasing very rapidly. Despite recent advancements, there still exist some unexplored areas such as, Zero-Shot Learning (ZSL) which aim’s to identify the embeddings of unseen entities. To deal with zero-shot learning we use both structural and textual descriptions of entities. For structural representation, we incorporate time directly into the vector space. For textual representation, we collect text descriptions of entities and use Convolutional Neural Networks (CNN) to capture the semantic features of the text descriptions. We utilize both of the aforementioned representations in the same vector space. Thus, we can utilize both structural and semantic representations of an entity which helps us to analyze the context of the entity. To test our model we utilized YAGO and Wikidata, two commonly used temporal knowledge graphs. We compare our method with two baseline methods. The evaluation result shows that our proposed method outperforms other methods for identifying relationships of unseen entities with proper time annotations.