Retrieving Relationships from a Knowledge Graph for Question Answering

Answering natural language questions posed on a knowledge graph requires traversing an appropriate sequence of relationships starting from the mentioned entities. To answer complex queries, we often need to traverse more than two relationships. Traditional approaches traverse at most two relationships, as well as typically first retrieve candidate sets of relationships using indexing etc., which are then compared via machine-learning. Such approaches rely on the textual labels of the relationships, rather than the structure of the knowledge graph. In this paper, we present a novel approach KG-REP that directly predicts the embeddings of the target relationships against a natural language query, avoiding the candidate retrieval step, using a sequence to sequence neural network. Our model takes into account the knowledge graph structure via novel entity and relationship embeddings. We release a new dataset containing complex queries on a public knowledge graph that typically require traversal of as many as four relationships to answer. We also present a new benchmark result on a public dataset for this problem.