Recurrent Neural Network Embedding for Knowledge-base Completion

Knowledge can often be represented using entities connected by relations. For example, the fact that tennis ball is round can be represented as “TennisBall HasShape Round”, where a “TennisBall” is one entity, “HasShape” is a relation and “Round” is another entity. A knowledge base is a way to store such structured information, a knowledge base stores triples of the “an entity-relation-an entity” form, and a real world knowledge base often has millions or billions of such triples. There are several well-known knowledge bases including FreeBase [1], WordNet [2], YAGO [3], etc. They are important in fields like reasoning and question answering; for instance if one asks “what is the shape of a tennis ball”, we can search the knowledge base for the triple as “TennisBall HasShape Round” and output “round” as the answer.