Modern Information Retrieval methods are capable of producing high-quality document rankings for a given search query. However, users of a retrieval system are often responsible for the selection and examination of retrieved documents and must read multiple documents to obtain all the information required for a full understanding of a topic. We seek to remove this responsibility from the user by instead constructing a single overarching document that incorporates information from all retrieved documents. Our multi-stage summarization approach avoids potential challenges that arise when summarizing large amounts of text by breaking the summarization problem into several subtasks. We first extract information from source documents, then eliminate redundancies, before finally ordering information in a logical manner. This approach is applied to the TREC-CAR Dataset, on which we demonstrate state-of-the-art results.
