MultiHumES: Multilingual Humanitarian Dataset for Extractive Summarization

Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics, pages 1713–1717 April 19 - 23, 2021. ©2021 Association for Computational Linguistics 1713 MultiHumES: Multilingual Humanitarian Response Dataset for Extractive Summarization Jenny Paola Yela-Bello EPFL Lausanne, Switzerland ypyelab@gmail.com Ewan Oglethorpe Data Friendly Space Virginia, United States ewan@datafriendlyspace.org Navid Rekabsaz Johannes Kepler University Linz, Austria navid.rekabsaz@jku.at Abstract When responding to a disaster, humanitarian experts must rapidly process large amounts of secondary data sources to derive situational awareness and guide decision-making. While these documents contain valuable information, manually processing them is extremely time- consuming when an expedient response is nec- essary. To improve this process, effective sum- marization models are a valuable tool for hu- manitarian response experts as they provide digestible overviews of essential information in secondary data. This paper focuses on ex- tractive summarization for the humanitarian re- sponse domain and describes and makes pub- lic a new multilingual data collection for this purpose. The collection – called MultiHumES – provides multilingual documents coupled with informative snippets that have been anno- tated by humanitarian analysts over the past four years. We report the performance re- sults of a recent neural networks-based sum- marization model together with other base- lines. We hope that the released data collection can further grow the research on multilingual extractive summarization in the humanitarian response domain. 1 Introduction The disaster risk management cycle consists of four stages: mitigation, preparedness, response, and re- covery (Alexander, 2002). The review of secondary data sources (i.e., reports, news, and other forms of text data) is embedded in all these stages, with vary- ing levels of importance from stage to stage. The work of secondary data review is characterized by a high and ever-increasing amount of information to be analyzed. At the same time, typically, only a small workforce is available to analyze such infor- mation. Early in the response phase, namely in the first 72 hours after a disaster strikes, secondary data review to gain situational awareness is essential, as it brings to light which type of relief activities to undertake. After this stage, primary data collection (such as surveys) begins while still supported by the secondary data review processes. The disaster information cycle by its part con- sists of the collection, collation, analysis, dissemi- nation, decision-making, and reporting stages. An effective summarization tool can provide meaning- ful support for the collection stage when analysts prioritize what documents to read first (i.e., offering an overview of a document). In the collation stage, when analysts need to take and merge the most im- portant findings from several documents, and even in the reporting stage, analysts are asked to bring in a few sentences containing the key findings of what they have written. An auto summarization system aims to provide analysts with a starting point from which they can continue their work rather than re- place it. Such a system can significantly save time in the overall disaster information cycle. In this work, we take the initial …

Content truncated to 500 words