CrumbTrail: An efficient methodology to reduce multiple inheritance in knowledge graphs

In this paper we present CrumbTrail, an algorithm to clean large and dense knowledge graphs. CrumbTrail removes cycles, out-of-domain nodes and non-essential nodes, i.e., those that can be safely removed without breaking the knowledge graph's connectivity. It achieves this through a bottom-up topological pruning on the basis of a set of input concepts that, for instance, a user can select in order to identify a domain of interest. Our technique can be applied to both noisy hypernymy graphs - typically generated by ontology learning algorithms as intermediate representations - as well as crowdsourced resources like Wikipedia, in order to obtain clean, domain-focused concept hierarchies. CrumbTrail overcomes the time and space complexity limitations of current state-of-art algorithms. In addition, we show in a variety of experiments that it also outperforms them in tasks such as pruning automatically acquired taxonomy graphs, and domain adaptation of the Wikipedia category graph.