Troll Tweet Detection Using Contextualized Word Representations
share
In recent years, numerous troll accounts that manipulate social media sentiment have emerged. Due to the use of fake and automated accounts by businesses, abusers, and nation-state-sponsored troll farms, detecting and eliminating trolling is a crucial issue for social networking platforms. Various NLP techniques are used to extract information from social networking text, such as tweets on Twitter, to identify the messages originating from fake accounts. In this context, this paper implements and compares nine deep learning-based architectures for troll tweet detection, with three models for each BERT, ELMo, and GloVe word embedding model. The majority of BERT-based architectures improve the detection of trolling tweets, as demonstrated by experimental results. An ELMo-based architecture with a GRU classifier has the highest AUC for detecting troll messages. Future socially-based systems can utilize the proposed architectures to identify troll messages.
READ FULL TEXT