Concept Extensions as the Basis for Vector-Space Semantics: Combining Distributional and Ontological Information about Entities

We propose to base the development of vector-space models of semantics on concept extensions, which defines concepts to be sets of entities. We investigate two sources of knowledge about entities that could be relevant: distributional information provided by word or phrase embeddings, and ontological information derived from a knowledge base. We develop a feedforward neural network architecture to learn entity representations that are used to predict their concept memberships, and show that the two sources of information are actually complementary. In an entity ranking experiment, the combination approach that uses both types of information outperforms models that only rely on one of the two. We also perform an analysis of the output using fuzzy logic techniques to demonstrate the potential of learning concept extensions for supporting inference involving classical semantic operations.