Answering Complex Questions Using Curated and Extracted Knowledge Bases

Abstract

The web is awash in textual data. As users struggle to navigate this textual data, the art and science of search engines have changed dramatically in recent years. Search engines like Google are now focusing on providing concise answers in response to user questions asked in natural language, instead of delivering an assortment of links to other websites. This has been made possible with the renaissance of large-scale knowledge bases (KBs), which contain facts about real-world entities and relations between them. Curated manually or extracted automatically from textual data, KBs have helped unlock the value in abundant unstructured textual sources. Despite the tremendous progress in question-answering over a knowledge base (KB-QA), existing systems still struggle to answer a wide variety of questions, especially complex questions where multiple pieces of information have to be combined to conclude the answer. This dissertation studies the design of KB-QA systems that can answer complex questions by reasoning over knowledge bases, curated manually or extracted automatically.

KB-QA systems face two challenges. The first challenge is the loss of information at knowledge acquisition: how do we prevent extraction systems from losing contextual information critical to answering complex questions. We describe open information techniques that are robust to complex textual data and can encode a complex fact as a set of linked simple facts. The extracted facts can be used to populate a KB automatically. The second challenge is querying: how do we translate complex questions to queries to access the information in the KB. The difficulty of this task and the coverage of the KB-QA system depend on the target KB. Automatically extracted KBs offer a high coverage of information but use many different patterns to express the information, making them hard to query. Manually curated KBs, on the other hand, suffer from low coverage due to their restricted schema, but can support compositional reasoning since they are constructed precisely for querying. We describe querying techniques for complex question-answering over each type of KB. We also describe a KB-QA system that can benefit from combining high-quality curated knowledge with broad-coverage automatically extracted facts for answering complex questions.