TrollsWithOpinion: A taxonomy and dataset for predicting domain-specific opinion manipulation in troll memes

Memes have become a de-facto media device in online communication. Unfortunately, memes are also used for trolling, which intends to demean, harass, or bully targeted individuals. As a result of which, the targeted individual could fall prey to opinion manipulation. Trolling via Image With Text (IWT) memes which we refer to as ‘troll memes’, are difficult to identify due to the multimodal (image + text) nature of such memes. However, the research into the identification and classification of troll memes with opinion manipulation remains unexplored. To bridge this research gap, we introduce a three-level taxonomy that studies the effect of trolling in domain-specific opinion manipulation. On the first level, we classify the meme as troll or not_troll. On the second level, we classify if the meme intends opinion manipulation. On the third level, if the opinion manipulation is present, then we classify the domain (political, product, other) of the opinion manipulation. To support the class definitions proposed in the taxonomy, we enhanced an existing dataset (Memotion) by annotating the data with our defined classes. This results in a dataset of 8,881 IWT memes in the English language (TrollsWithOpinion dataset) which we make available as open-source at Github(https://github.com/sharduls007/TrollOpinionMemes). We perform experiments on all three levels and present the classification report of the results using Machine Learning and state-of-the-art Deep Learning techniques. The classification report highlights the complex nature of the task since the models perform well on the first two levels. However, we see a degradation of the evaluation results on the third level of the taxonomy.