Measuring Triplet Trustworthiness in Knowledge Graphs via Expanded Relation Detection.

Nowadays, large scale knowledge graphs are usually constructed by (semi-)automatic information extraction methods. Nevertheless, the technology is not perfect, because it cannot avoid introducing erroneous triplets into knowledge graphs. As a result, it is necessary to carry out some screening of the trustworthiness of the triplets in knowledge graphs before putting them into industrial use. In this paper, we propose a novel framework named as KGerd, for measuring triplet trustworthiness via expanded relation detection. Given a triplet (h, r, t), we center our framework on the basis of the classic translation-based mechanism among h, r and t. Besides translation-based relation detection, we introduce two additional types of relation detection approaches, which consider, respectively, to expand the task vertically by leveraging abstract versions of the relation r, as well as laterally by generating connecting paths between the entities h and t. The three detection results are then combined to provide a trustworthiness score for decision making. Comprehensive experiments on real-life datasets demonstrate that our proposed model KGerd is able to offer better performance over its state-of-the-art competitors.