Weakly-Supervised Fine-Grained Event Recognition on Social Media Texts for Disaster Management

Abstract

People increasingly use social media to report emergencies, seek help or share information during disasters, which makes social networks an important tool for disaster management. To meet these time-critical needs, we present a weakly supervised approach for rapidly building high-quality classifiers that label each individual Twitter message with fine-grained event categories. Most importantly, we propose a novel method to create high-quality labeled data in a timely manner that automatically clusters tweets containing an event keyword and asks a domain expert to disambiguate event word senses and label clusters quickly. In addition, to process extremely noisy and often rather short user-generated messages, we enrich tweet representations using preceding context tweets and reply tweets in building event recognition classifiers. The evaluation on two hurricanes, Harvey and Florence, shows that using only 1-2 person-hours of human supervision, the rapidly trained weakly supervised classifiers outperform supervised classifiers trained using more than ten thousand annotated tweets created in over 50 person-hours.

Semiautomated social media analytics for sensing societal impacts due to community disruptions during disasters

“…The upper part of Figure 1 visualizes the framework of the proposed method of social media analytics for retrieving disruptions of critical infrastructures and societal impacts. SocialDISC labels disaster‐related social media posts according to predefined taxonomy (Table 1) indicating different types of community disruptions and filters out irrelevant posts (Yao, Zhang, Saravanan, Huang, & Mostafavi, 2020). Within each category in the taxonomy, a content distilling process enables the analyzers to determine the topic of each subevent efficiently by scanning the situational information posts and the keywords with high frequency.…”

“…In the content distilling process, a network‐based clustering algorithm first clusters social media posts in each category into clusters according to content similarity, enabling users to read similar social media posts more efficiently. Within each information cluster, a pretrained classifier (Yao et al., 2020) then separates social media posts into situational information (posts describing the situation using a formal and objective language) or residents’ reaction (posts expressing personal feelings, comments, and complaints about the situation using a casual and subjective language). Finally, SocialDISC will assess emotion scores of the societal reaction, a quantitative indicator of the societal impact caused by the disruption event for each subevent.…”

“…In phase three, we iteratively retrain the multichannel classifier to improve its performance of event recognition. To see more technical details about the weakly supervised labeling algorithm, specifically for the presentation of the WSD algorithm and the BiLSTM classifier, please refer to another work from the authors (Yao et al., 2020).…”

“…Furthermore, domain adaptation of supervised classification is challenging overall, which exacerbates the limitation of supervised learning caused by the paucity of training data. To increase the adaptivity and timeliness, SocialDISC uses a weakly supervised approach to label social media posts (Yao et al, 2020). This labeling approach mainly consists of three main phases ( Figure 2).…”

“…In phase one, we applied automatic clustering of social media posts (i.e., tweets in this study) containing preliminary keywords for each event category (based on the taxonomy in Table 1) and then conducted manual word sense disambiguation (WSD) on the clusters to generate training instances for the classifier. In phase two, we trained a F I G U R E 2 Overview of the weakly supervised labeling of social media posts (Yao et al, 2020) multichannel Bidirectional Long Short-Term Memory (BiL-STM) classifier (an artificial recurrent neural network architecture used in the field of deep learning; Graves & Schmidhuber, 2005) using tweets together with their context tweets and reply tweets. We chose the BiLSTM model because it is proved to be the state-of-the-art in classifying natural language data and capturing the semantics of sentences (T. Chen, Xu, He, & Wang, 2017;Peters et al, 2018;Zhou et al, 2016).…”