Useful event detection on online social medias

Abstract

Social media has transformed the Internet into a dynamic platform, enabling users to create, share, and comment on textual content related to various categories of events. However, processing data derived from social media presents a significant challenges for event detection and classification due to the massive volume of information and the often informal structure of texts. To address this issue, research has highlighted the importance of corpus and datasets, which are indispensable for analyzing and processing textual con- tent from social media, particularly for event (ED) and classification. In this thesis, we propose solutions for event detection and classification, with a specific focus ontweets published on the social media platform Twitter. Event detection and classification are especially crucial in numerous fields, such as politics, healthcare, disaster manage- ment, science, sports, economics and others. Our contributions include the construction of datasets and the development of advanced models for detection and classification, leveraging recent advancements in deep learning, notably transformer architectures such as BERT model. Specifically, a variety of model variants have been experimented toaddress different contexts and requirements, these include BERT Base, BERT Large, DistilBERT, CAMeLBERT, ARAELECTRA, among others. Our experiments with these models have shown promising results compared to the state of the art, achieving impres- sive accuracy rates that have reached more than 94% in some cases. These out comes underscore the relevance of our approaches and their potential applications across various domains, specifically in event detection and classification of text in OSM.