Augmenting knowledge graphs with natural language evidence

Abstract

In last ten years, massive amounts of world knowledge have been accumulated into large knowledge graphs (KGs). These knowledge repositories store millions of facts about the world, such as information about people, places and organizations, and have become a powerful asset for semantic applications such as search, analytics, recommendations, and data integration. Several approaches have been proposed to create KGs from Wikipedia, as in the cases of YAGO and DBpedia, or collaboratively as for Freebase and Wikidata. Despite their seemingly huge size, these knowledge graphs are greatly incomplete and approaches to populate them automatically are needed to increase their coverage. This thesis describes principled methods to model knowledge graph relations with natural language. These models allow the extraction of facts from text or to annotate web tables with KG relations, with the aim of populating state-of-the-art KGs. The first contribution is a pattern-based extraction system which can extract automatically highquality facts from the text of Wikipedia articles. Indeed, the approaches used to derive KGs from Wikipedia are focused only on its structured components like the info-boxes. Although valuable, they represent only a fraction of the actual information expressed in the articles. We experiment our system on five different languages, showing that it can extract a large number of facts that are out of reach of common infobox-based extractions. The second contribution is an approach that uses language models, derived from aWeb-scale corpus, to rank KG relations that hold over pairs of entities juxtaposed in tables or structured lists. Our experimental evaluation shows the effectiveness of the approach in predicting KG relations even when entities are missing from the graph and thus represents a significant advancement of the state-of-the-art.