proBE–A BERTweet-Attentive Prototype Few-Shot Model for Event Detection in OSN

In the domain of digital communication, a significant transition have occurred, with social networks overtaking traditional news media outlets to become the primary source of information on a global scale. This evolution has given rise to a dynamic environment, where user-generated content escalates at an exceptional pace. However, managing this vast and ever-changing data poses significant challenges. Annotating such voluminous and dynamic information is not only financially burdensome but also impractical in terms of feasibility. In addition, current learning techniques face significant challenges when it comes to identifying unseen or novel occurrences in social network data, while also demanding a huge amount of training data to operate effectively. This entails a new era that emphasis on the need for less data and more generalisation, similar to how humans interpret information. To this end, in this research, we developed the model proBE, which formalises event detection in online social networks as a few-shot learning problem and offers a novel perspective on it. The suggested method encodes the tweet messages (aka tweets) with BERTweet, to capture context with respect to inbuilt features of Twitter like hashtags, emoticons and then employed an attentive prototype model where, tweet attention and feature attention is applied to highlight the contextually rich key tweets and the prominent features, respectively. proBE is evaluated on real world benchmark Twitter datasets CrisisLexT26 and CrisisLexT6 and performs significantly better when compared to various baseline methods in terms of accuracy, F-score, precision and recall. To the best of our knowledge, this is the first study that employs few-shot learning in event detection in online social networks to overcome the sparsity of labelled data due to the volatile dynamics of online social networks and the inability of existing approaches to detect unseen events even after enormous amount of training data.