Integration of New Concepts in Knowledge Hierarchies
12 July 2018
Knowledge hierarchies have become immensely useful and popular in a wide variety of fields owing to their applicability in structurally organizing and conveniently navigating information. A plethora of techniques have been developed to automatically build knowledge hierarchies from a variety of data sources. Nevertheless, maintaining such extensive knowledge bases and frequently updating them with current knowledge advances is a considerable challenge. In this work, we propose a novel technique to extend existing large, generic knowledge hierarchies with new, emergent concepts from up-to-date resources such as news, social media and research publications. For each existing as well as new concept, we learn a high-dimensional vector embedding via a generated context of terms, followed by identifying the neighbors of the new concepts in the embedding space. We then predict the potential parents of the new concepts in the knowledge hierarchy by analyzing the ancestors of their neighbors, and ranking them based on a set of graph features. We evaluated our approach on the DBPedia and WordNet concept hierarchies, and were able to achieve an NDCG score greater than 70% and an F1-score of up to 83%, significantly improving over baselines. We finally perform a case study on emergent, real-world concepts from specific domains and achieve promising performance.
