Predicting Salient Updates for Disaster Summarization

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 1608–1617, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics Predicting Salient Updates for Disaster Summarization Chris Kedzie and Kathleen McKeown Columbia University Department of Computer Science {kedzie, kathy}@cs.columbia.edu Fernando Diaz Microsoft Research fdiaz@microsoft.com Abstract During crises such as natural disasters or other human tragedies, information needs of both civilians and responders often re- quire urgent, specialized treatment. Moni- toring and summarizing a text stream dur- ing such an event remains a difficult prob- lem. We present a system for update sum- marization which predicts the salience of sentences with respect to an event and then uses these predictions to directly bias a clustering algorithm for sentence se- lection, increasing the quality of the up- dates. We use novel, disaster-specific features for salience prediction, including geo-locations and language models repre- senting the language of disaster. Our eval- uation on a standard set of retrospective events using ROUGE shows that salience prediction provides a significant improve- ment over other approaches. 1 Introduction During crises, information is critical for first re- sponders, crisis management organizations, and those caught in the event. When the event is sig- nificant, as in the case of Hurricane Sandy, the amount of content produced by traditional news outlets, government agencies, relief organizations, and social media can vastly overwhelm those try- ing to monitor the situation. Crisis informatics (Palen et al., 2010) is dedicated to finding methods for sharing the right information in a timely fash- ion during such an event. Research in this field has focused on human-in-the-loop approaches rang- ing from on the ground information gathering to crowdsourced reporting and disaster management (Starbird and Palen, 2013). Multi-document summarization has the poten- tial to assist the crisis informatics community. Au- tomatic summarization could deliver relevant and salient information at regular intervals, even when human volunteers are unable to. Perhaps more im- portantly it could help filter out unnecessary and irrelevant detail when the volume of incoming in- formation is large. While methods for identifying, tracking, and summarizing events from text based input have been explored extensively (Allan et al., 1998; Filatova and Hatzivassiloglou, 2004; Wang et al., 2011), these experiments were not devel- oped to handle streaming data from a heteroge- neous environment at web scale. These methods also rely heavily on redundancy which is subop- timal for time sensitive domains where there is a high cost in delaying information. In this paper, we present an update summariza- tion system to track events across time. Our sys- tem predicts sentence salience in the context of a large-scale event, such as a disaster, and integrates these predictions into a clustering based multi- document summarization system. We demonstrate that combining salience with clustering produces more relevant summaries compared to baselines using clustering or relevance alone. Our experi- ments suggest that this is because our system is better able to adapt to dynamic changes in input …

Content truncated to 500 words