Methods and Tools for Summarization of Entities and Facts in Knowledge Bases

Abstract

Knowledge bases have become key assets for search and analytics over large document corpora. They are used in applications ranging from highly specialized tasks in bioinformatics to general purpose search engines. The large amount of structured knowledge they contain calls for effective summarization and ranking methods. The goal of this dissertation is to develop methods for automatic summarization of entities in knowledge bases, which also involves augmenting them with information about the importance of particular facts on entities of interest. We make two main contributions. First, we develop a method to generate a summary of information about an entity using the type information contained in a knowledge base. We call such a summary a semantic snippet. Our method relies on having importance information about types, which is external to the knowledge base. We show that such information can be obtained using human computing methods, such as Amazon Mechanical Turk, or extracted from the edit history of encyclopedic articles in Wikipedia. Our second contribution is linking facts to their occurrences in supplementary documents. Information retrieval on text uses the frequency of terms in a document to judge their importance. Such an approach, while natural, is difficult for facts extracted from text. This is because information extraction is only concerned with finding any occurrence of a fact. To overcome this limitation we propose linking known facts with all their occurrences in a process we call fact spotting. We develop two solutions to this problem and evaluate them on a real world corpus of biographical documents.