Detecting Perceived Emotions in Hurricane Disasters

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5290–5305 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 5290 Detecting Perceived Emotions in Hurricane Disasters Shrey Desai1 Cornelia Caragea2 Junyi Jessy Li1 1The University of Texas at Austin 2University of Illinois at Chicago {shreydesai@, jessy@austin.}utexas.edu cornelia@uic.edu Abstract Natural disasters (e.g., hurricanes) affect mil- lions of people each year, causing widespread destruction in their wake. People have recently taken to social media websites (e.g., Twitter) to share their sentiments and feelings with the larger community. Consequently, these platforms have become instrumental in under- standing and perceiving emotions at scale. In this paper, we introduce HURRICANEEMO, an emotion dataset of 15,000 English tweets spanning three hurricanes: Harvey, Irma, and Maria. We present a comprehensive study of fine-grained emotions and propose classi- fication tasks to discriminate between coarse- grained emotion groups. Our best BERT (De- vlin et al., 2019) model, even after task-guided pre-training which leverages unlabeled Twitter data, achieves only 68% accuracy (averaged across all groups). HURRICANEEMO serves not only as a challenging benchmark for mod- els but also as a valuable resource for analyz- ing emotions in disaster-centric domains. 1 Introduction Natural disasters cause thousands of deaths and dis- place hundreds of millions each year (Ritchie and Roser, 2020). These catastrophic events not only induce material destruction but also stir an integral part of being human: our emotions. Disasters ad- versely affect individuals’ mental states (Fritz and Marks, 1954; Kinston and Rosser, 1974), and there- fore it is no surprise that many take to social media (e.g., Twitter) to share their feelings. Social me- dia websites, as a result, have become an essential platform for understanding the expression and per- ception of emotions at a significantly larger scale (Mohammad, 2012; Wang et al., 2012; Moham- mad and Kiritchenko, 2015; Volkova and Bachrach, 2016; Abdul-Mageed and Ungar, 2017), with far reaching potential influences from academic re- search to public policy (Dennis et al., 2006; Fritze et al., 2008; Fraustino et al., 2012). While natural language processing methods have been effective for emotion detection (Strapparava and Mihalcea, 2007), existing resources struggle in disaster-centric domains, in part due to distribu- tional shifts. Emotion detection in natural disasters (e.g., hurricanes) requires implicit reasoning not available as surface-level lexical information. For example, in “of course, [we]1 still have the [storm surge]2 coming,” given the context, we can rea- sonably infer discontent towards the “storm surge” despite the absence of polarizing words. There- fore, distantly supervised techniques largely based on lexical units (Mohammad and Turney, 2013; Abdul-Mageed and Ungar, 2017) fail to capture this type of deeper semantic phenomena. Our paper presents a comprehensive investiga- tion into perceived emotions in hurricane disas- ters. To this end, we introduce HURRICANEEMO, a dataset of 15,000 disaster-related tweets (in En- glish) streamed during Hurricanes Harvey, Irma, and Maria, which were devastating tropical storms occurring in the 2017 Atlantic hurricane season (Belles, 2017). Our samples are annotated with fine-grained emotions derived from the Plutchik Wheel …

Content truncated to 500 words