Reading and Reasoning with Knowledge Graphs

Much attention has recently been given to the creation of large knowledge bases that contain millions of facts about people, things, and places in the world. These knowledge bases have proven to be incredibly useful for enriching search results, answering factoid questions, and training semantic parsers and relation extractors. The way the knowledge base is actually used in these systems, however, is somewhat shallow—they are treated most often as simple lookup tables, a place to find a factoid answer given a structured query, or to determine whether a sentence should be a positive or negative training example for a relation extraction model. Very little is done in the way of reasoning with these knowledge bases or using them to improve machine reading. This is because typical probabilistic reasoning systems do not scale well to collections of facts as large as modern knowledge bases, and because it is difficult to incorporate information from a knowledge base into typical natural language processing models. 

In this thesis we present methods for reasoning over very large knowledge bases, and we show how to apply these methods to models of machine reading. The approaches we present view the knowledge base as a graph and extract characteristics of that graph to construct a feature matrix for use in machine learning models. The graph characteristics that we extract correspond to Horn clauses and other logic statements over knowledge base predicates and entities, and thus our methods have strong ties to prior work on logical inference. We show through experiments in knowledge base completion, relation extraction, and question answering that our methods can successfully incorporate knowledge base information into machine learning models of natural language.