Explaining completions produced by embeddings of knowledge graphs

© Springer Nature Switzerland AG 2019. Advanced question answering typically employs large-scale knowledge bases such as DBpedia or Freebase, and are often based on mappings from entities to real-valued vectors. These mappings, called embeddings, are accurate but very hard to explain to a human subject. Although interpretability has become a central concern in machine learning, the literature so far has focused on non-relational classifiers (such as deep neural networks); embeddings, however, require a whole range of different approaches. In this paper, we describe a combination of symbolic and quantitative processes that explain, using sequences of predicates, completions generated by embeddings.