An approach to detect offence in Memes using Natural Language Processing(NLP) and Deep learning

Social media is one of the most popular form of platform which is most common with people of our age. With passage of time, memes have gained a significant popularity and are often shared on social media platforms. Memes usually have hilarious content but can be offensive sometimes, containing some hateful message or character image. Such memes may have detrimental social impact in our society or to any individual. [1] Thus, an automated system for evaluation of offensiveness in the meme content is required. This paper presents an approach to detect offense in memes using Natural Language Processing (NLP) and deep learning. Due to the increasing number of memes over the internet, it is not an easy task to evaluate each meme before it spreads all around. This raises a demand for a system that can automate the process of evaluating memes before they agitate a crowd or spread a humour. This paper presents a model to detect offensive memes, in three steps. First, it will extract the text from the given image, then it will classify the given text as offensive or not offensive. If the text is found to be offensive then in the third step it will further classify offensive text in three categories namely slight offensive, very offensive and hateful offensive. The dataset used for this work consists of 6,992 memes which were labeled as not offensive, slightly offensive, very offensive, and hateful offensive. The model uses very simple architecture with a multi-layer dense network structure involving NLP with RNN and LSTM along with word embeddings such as GloVe and FastText.