Distilling Causal Metaknowledge from Massive Knowledge Graph

Abstract

In recent years, the growing information overload facilitates the access to billions of relational facts in the world, which are usually integrated in all manner of knowledge graphs. The metaknowledge, defined as the knowledge about knowledge, reveals the inner principle of arising these factual knowledge, and hence is of vital importance to be discovered for the understanding, exploiting and completion of knowledge. In this paper, we focus on capturing the causal component of metaknowledge, that is a metarule with causal semantic. For the propose, we devise an efficient causal rule discovery algorithm called CaRules that distills the causal rules between two knowledge graph schemata abstracted from instances from massive knowledge graphs. Extensive experiments demonstrate that the quality and interpretability of the causation-based rules outperform the correlation-based rules, especially in the out-of-distribution tasks.

Tasks

Reproductions