An improved textual storyline generating framework for disaster information management

Abstract

Analyzing and understanding disaster-related sit­uation updates from a large number of disaster-related docu­ments plays an important role in disaster management and has attracted a lot of research attention. Recently several methods have been developed to generate textual storylines from disaster-related documents to help people understand the overall trend and evolution of a disaster event as well as how the disaster affects different areas. These methods are able to help people improve their situation awareness by generating informative summarizes to present the global pictures of disaster events. However, these methods suffer from several limitations including text representation, repre­sentative document selection, and summary generation that may affect the quality of the summarized results. To address these limitations, in this paper, we propose an improved two-layer storyline generating framework which generates a global storyline of the disaster events in the first layer, and provides condensed information about specific regions affected by the disaster in the second layer. The proposed framework utilizes the word embedding for text similarity measurement, considers both uniqueness and relevance for representative document selection, and uses Maximal Marginal Relevance to generate summaries from each local document set. The experimental results on four typhoons related events demonstrate the efficacy of our proposed framework on capturing the status information and understanding the situation from a large of documents.