Detecting Perceived Emotions in Hurricane Disasters

Abstract

Natural disasters (e.g., hurricanes) affect millions 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 understanding 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 classification tasks to discriminate between coarsegrained emotion groups. Our best BERT (Devlin 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 models but also as a valuable resource for analyzing emotions in disaster-centric domains. Related WorkEmotion detection has been extensively studied in news headlines (Strapparava and Mihalcea, 2007;

An analysis of emotions and the prominence of positivity in #BlackLivesMatter tweets

“…For example, the National Research Council Canada (NRC) Word-Emotion Association Lexicon (EmoLex), which contains words associated with eight emotions, associates “police” with fear, positive, and trust, which are contradictory to the connotations of “police” in protests against police brutality ( 15 , 25 ). More recently, machine learning–based NLP models have outperformed traditional lexicon approaches at identifying affect in text ( 18 , 19 , 26 , 27 ). Neural models are trained on annotated datasets and used to infer affect in unseen text.…”

“…The same model can then be fine-tuned for a specific task. Following prior work, we use masked language model pretraining over unannotated sentences from the target data to encourage domain adaption and then, fine-tune the model to infer emotions using the annotated source data, as in BASE ( 19 , 29 , 30 ).…”

“…Previous examinations of emotions and affect expressed in tweets about the Black Lives Matter movement have relied on lexicon (Linguistic Inquiry and Word Count, LIWC) scores ( 3 ), and analyses of other protest events have similarly relied on lexicon-based approaches ( 15 , 16 ). While recent research has led to the development of more powerful deep learning–based models and annotated datasets, these models nevertheless are prone to overfitting to shallow lexical cues and often perform poorly in new domains ( 17 – 19 ). Thus, in this work, we leverage recent natural language processing (NLP) techniques, including in-domain pretraining and few-shot learning, to improve emotion analysis model performance across domains in an easily adaptable framework.…”

CancerEmo: A Dataset for Fine-Grained Emotion Detection

“…Emotion detection has been studied in computational linguistics for a long time, with researchers exploring domains ranging from music and classic literature (Liu et al, 2019a) to social networks (Mohammad, 2012;Islam et al, 2019;Desai et al, 2020) and online news (Bao et al, 2009). Most studies focus on two main emotion categorizations: Ekman's (Ekman, 1992) 6 basic emotions (Katz et al, 2007;Aman and Szpakowicz, 2007;Mohammad, 2012) and Plutchik's (Plutchik, 1980) 8 emotions (Abdul-Mageed and Ungar, 2017;Mohammad and Turney, 2010).…”

eMLM: A New Pre-training Objective for Emotion Related Tasks

“…There are numerous studies that focus on emotion detection (Demszky et al, 2020;Desai et al, 2020;del Arco et al, 2020;Sosea and Caragea, 2020;Majumder et al, 2019;Mohammad and Kiritchenko, 2018;Abdul-Mageed and Ungar, 2017;Mohammad and Kiritchenko, 2015;Mohammad, 2012;Strapparava and Mihalcea, 2008) and sentiment analysis (Yin et al, 2020;Tian et al, 2020;Phan and Ogunbona, 2020;Zhai and Zhang, 2016;Chen et al, 2016;Liu, 2012;Glorot et al, 2011;Pang and Lee, 2005). Various lexicons have been used to improve model performance on these tasks.…”