Unimodal, multimodal and transformational approaches for offensive meme classification: challenges, datasets, and models

The proliferation of offensive content in online platforms presents a critical challenge for maintaining a respectful and inclusive digital environment. This thesis addresses the problem of offensive meme classification through a multimodal approach that leverages textual and visual cues. The overarching goal is to analyze the challenges and improve the state-of-the-art detecting and categorising offensive memes in diverse forms, spanning text, images, and their combinations. We introduce novel datasets, MultiOFF, TrollsWithOpinion, and TamilMemes, which encompass a diverse array of offensive content instances in both textual and image formats. These datasets facilitate the training and evaluation of our multimodal classification models. The contributions of this thesis extend beyond the development of effective offensive meme classification datasets and models. We present a comprehensive analysis of the challenges and opportunities associated with multimodal offensive meme detection. Furthermore, we release all the datasets and the developed models as open-source resources to support future research in this critical domain. In this thesis, we answer three research questions. The first defines the domain-specific opinion manipulation by offensive memes, and proposes a taxonomy to categorize memes. The second identifies the effectiveness of the methods of developing datasets for the multimodal meme classification task. Lastly, the third one tests the effectiveness of the unimodal, multimodal, and transformational methods for the offensive meme classification task.

Publisher

University of Galway

Rights

Attribution-NonCommercial-NoDerivatives 4.0 International

cbn

Collections

University of Galway Theses (PhD Theses)