Multi-task Knowledge Graph Representations via Residual Functions

In this paper, we propose MuTATE, a Multi-Task Augmented approach to learn Transferable Embeddings of knowledge graphs. Previous knowledge graph representation techniques either employ task-agnostic geometric hypotheses to learn informative node embeddings or integrate task-specific learning objectives like attribute prediction. In contrast, our framework unifies multiple co-dependent learning objectives with knowledge graph enrichment. We define co-dependence as multiple tasks that extract covariant distributions of entities and their relationships for prediction or regression objectives. We facilitate knowledge transfer in this setting: tasks -> graph, graph -> tasks, and task-1 -> task-2 via task-specific residual functions to specialize the node embeddings for each task, motivated by domain-shift theory. We show 5% relative gains over state-of-the-art knowledge graph embedding baselines on two public multi-task datasets and show significant potential for cross-task learning.