III: Small: Geometric Constraint based Concept Keyword Embedding for Domain-neutral Knowledge Graph Construction

Indiana University, Bloomington IN

Investigators

Abstract

Many knowledge-driven applications, including question answering systems and AI based intelligent personal assistants depend critically on the availability of high quality knowledge graphs. However, existing knowledge graphs suffer from many limitations, including lack of completeness, and excessive need for human intervention. The objective of this proposal is to overcome these limitations by learning meaningful representation vectors for the entities and the relationships by combining sentential context of entities from text documents with the relational context of entities from existing knowledge bases. Another objective is to develop methodologies for automatic extraction of entity-pairs and their relationships from natural language text. The main intellectual merit of the proposed methodologies is the design of representation learning methodologies which can impose geometric constraints on the embedding vectors for capturing the underlying semantics; in other words, using the proposed method, the disposition of the learned vectors would capture various semantic relationships among the entities, including hierarchical, compositional, and sequential, which would yield high quality representation vectors for the entities, leading to high-quality knowledge- graphs. Validation of the proposed methods will be performed by building a living knowledge graph of all of computer science. In terms of broader impact, the proposed methods would benefit any information-heavy disciplines, for which, the construct and temporal evolution of a knowledge graph are critical. Also, there is a direct benefit to computer science as a discipline, since organizing and summarizing all the known and emerging concepts in computer science within a living and encyclopedic knowledge graph will enable better organization of new knowledge, and better search of related works from the existing literature. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.