Towards Domain-Independent Multi-Lingual-Dialectal Online Social Behavior Modeling

Abstract

Online social networks (OSNs) have changed the way humans communicate. They seem to have transferred their real-life means of communication and their social behaviors to digital forms on the virtual social media. With this move to the "new world", they have not only adapted some of the already existing forms of communication to fit the new milieu, but they have also adopted new forms of communication provided by OSNs. With communications being shifted increasingly to OSNs, especially after the outbreak of COVID-19 pandemic, the need for tracking and understanding human behaviors online has risen. Also discovering emerging trends and concerns in order to understand the corresponding online social behavior (OSB) that best reflects its offline settings has become a necessity. Further, voicing out concerns and communicating timely trends are not restricted to a single spoken language; Facebook alone reported that two-third of its users speak languages other than English. Besides, the informal and slang nature of conversations and communication has become the new norm on social media platforms which, in turn, has triggered the need to understand foreign languages and even their dialects in order to be able to widely monitor OSB within countries and across the world. This is particularly vital to ensure stability and well-being in societies and to enhance the quality of decision-making in smart cities. Despite all those challenges, we have been able to analyze the users' OSBs, based on the users' textual and visual forms of communication as a first step. This does not involve literal translation, i.e. rendering a text from one language to another without considering the sense of the text, but rather includes examining the geo-cultural contextualization of OSN communications. This is significant as behavior is the outcome of a culture and is manifested through word use (language/dialect in this case). In response, we propose a multimedia framework for modeling domain-independent OSB in different languages and dialects used on OSNs. Unsupervised and supervised learning approaches have been utilized in developing the components of the proposed framework. The first component refers to content-localization based machine translation and is responsible for capturing the multi-lingual multi-dialectal aspect of OSN conversations using AI power. The second component is responsible for modeling textual and visual OSB using machine learning and deep learning algorithms. The third component presents topic modeling and dynamic topic interpretation and is responsible for inferring hidden patterns from a stream of multi-lingual-dialectal data and providing comprehensible interpretations as a step towards facilitating the analysis of the predicted OSB. Further, new datasets have been proposed and constructed to develop and evaluate our proposed AI-models. In addition to our comprehensive experimentations conducted to evaluate the proposed framework, our large-scale analysis of COVID-19 pandemic has reinforced the capability of our proposed solution to recognize concerns and trends, along with reliability to analyze multi-lingual-dialectal OSBs, using real OSN data collected from North America and Middle East regions. This thesis presents a comparative analysis of OSB and data discoveries in Canada, USA, Lebanon and Saudi Arabia.