Multidocument Summarization via Information Extraction

Multidocument Summarization via Information Extraction Michael White and Tanya Korelsky CoGenTex, Inc. Ithaca, NY mike,tanya@cogentex.com Claire Cardie, Vincent Ng, David Pierce, and Kiri Wagstaff Department of Computer Science Cornell University, Ithaca, NY cardie,yung,pierce,wkiri@cs.cornell.edu ABSTRACT We present and evaluate the initial version of RIPTIDES, a system that combines information extraction, extraction-based summarization, and natural language generation to support user- directed multidocument summarization. 1. INTRODUCTION Although recent years has seen increased and successful research efforts in the areas of single-document summarization, multidocument summarization, and information extraction, very few investigations have explored the potential of merging summarization and information extraction techniques. This paper presents and evaluates the initial version of RIPTIDES, a system that combines information extraction (IE), extraction-based summarization, and natural language generation to support user- directed multidocument summarization. (RIPTIDES stands for RapIdly Portable Translingual Information extraction and interactive multiDocumEnt Summarization.) Following [10], we hypothesize that IE-supported summarization will enable the generation of more accurate and targeted summaries in specific domains than is possible with current domain-independent techniques. In the sections below, we describe the initial implementation and evaluation of the RIPTIDES IE-supported summarization system. We conclude with a brief discussion of related and ongoing work. 2. SYSTEM DESIGN Figure 1 depicts the IE-supported summarization system. The system first requires that the user select (1) a set of documents in which to search for information, and (2) one or more scenario templates (extraction domains) to activate. The user optionally provides filters and preferences on the scenario template slots, specifying what information s/he wants to be reported in the summary. RIPTIDES next applies its Information Extraction subsystem to generate a database of extracted events for the selected domain and then invokes the Summarizer to generate a natural language summary of the extracted information subject to the user’s constraints. In the subsections below, we describe the IE system and the Summarizer in turn. 2.1 IE System The domain for the initial IE-supported summarization system and its evaluation is natural disasters. Very briefly, a top-level natural disasters scenario template contains: document-level information (e.g. docno, date-time); zero or more agent elements denoting each person, group, and organization in the text; and zero or more disaster elements. Agent elements encode standard information for named entities (e.g. name, position, geo-political unit). For the most part, disaster elements also contain standard event-related fields (e.g. type, number, date, time, location, damage sub-elements). The final product of the RIPTIDES system, however, is not a set of scenario templates, but a user-directed multidocument summary. This difference in goals influences a number of template design issues. First, disaster elements must distinguish different reports or views of the same event from multiple sources. As a result, the system creates a separate disaster event for each such account. Disaster elements should also include the reporting agent, date, time, and location whenever possible. In addition, damage elements (i.e. human and physical effects) are best grouped according to the reporting event. Finally, a slight broadening of the IE task was necessary in that extracted text was not …

Content truncated to 500 words