Abstract
With communications being shifted to online social networks (OSNs) as a result of travel and social restrictions during COVID-19 pandemic, the need has arisen for discovering emerging trends and concerns formed during the pandemic as well as understanding the corresponding online social behavior that reflects its offline settings. The online connectivity of devices through social media is one example of Internet of Things (IoT) in which a two-way communication between societies and officials, could be created. Therefore, it is possible to monitor people’s behavior through OSNs, especially during pandemics, to prevent potential social and psychological instabilities that might lead to undesired consequences. This is particularly crucial for governmental and non-governmental organizations to ensure the stability and well-being in societies. In response, we propose a pandemic-friendly real-time framework for monitoring cyber social behavior by utilizing unsupervised and supervised learning approaches. Two BERT-based supervised classifiers are trained and constructed to analyze two types of online social behaviors, hate and sentiment. Unsupervised framework is proposed for OSNs data exploration and coherent interpretation that is used as a complementary tool to facilitate the analysis of online social behaviors during pandemics. Extensive experimentation and evaluation have been conducted to validate the effectiveness of the proposed work. Our results have shown superior performance of our BERT-based models in two classification tasks: 1) binary classification for hate behavior detection and 2) multi-class classification for sentiment behavior detection. In addition to our experimentation results, our large-scale analysis of COVID-19 pandemic has illustrated the capability of our unsupervised framework for concerns and trends discoveries using OSNs data, along with reliability in automatically and dynamically providing phrase-based interpenetration of the inferred trends and concerns. This paper provides a twelve-month comparison analysis of data discoveries and online social behavior between Canada and USA during COVID-19 pandemic.
Highlights
With the COVID-19 global pandemic, the whole world has gone into an abrupt shift where every aspect of our lives has been impacted
We have found that BERT-based models are less sensitive to class imbalance when compared to LSTM and biLSTM models
We propose using topic modeling and phrase extraction methods for discovering hidden patterns and inferring topics, trends, and concerns formed during the pandemic as well as automatically providing coherent interpretation of the inferred topics without human effort involved
Summary
With the COVID-19 global pandemic, the whole world has gone into an abrupt shift where every aspect of our lives has been impacted. The challenge of modeling domain-dependent hate language sheds light on the importance of learning general patterns of hate language This general knowledge of models helps in capturing a wide spectrum of hate behavior across social media, and in controlling and detecting the spread of hate contents regardless of their types. This work is an attempt to assist in catering to public safety and psycho-social needs towards providing measures for developing healthy coping strategies to reduce the psycho-social instabilities during and post pandemic It could create opportunities for tracing individuals or groups responsible for violent incitement as it has been proven that it is possible to infer this type of information through OSNs [17], [18].
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
