Abstract

The use of vectors for representing the entities from a given knowledge base is now standard practice in Natural Language Processing. From a knowledge representation point of view, however, it seems more intuitive to model categories as regions (or distributions) rather than single vectors. A given individual is then assumed to belong to some category if the vector representation of that individual belongs to the corresponding region. Apart from increasing the interpretability of vector space representations, such region based representations also significantly expand the range of knowledge that can be expressed. In particular, we can show that region based vector space representations are able to capture a large sub-fragment of the class of existential rules. Unfortunately, estimating meaningful region representations in high-dimensional vector spaces is challenging, especially because they often have to be estimated from a very small number of examples. In our recent work, we have proposed a number of solutions that try to alleviate the lack of sufficient training examples by exploiting prior knowledge about the semantic relationships between different categories.