dhyani15/hatefulmemes

Contains code for Project - Multimodal Classification To Detect Hate Speech

As part of the project, we have designed, developed and trained a multimodal classification model on the Hateful Memes dataset. The challenge was initially organized by Facebook AI to detect hate speech in multimodal memes. The winner of the challenge had an accuracy of 84.50%. The state-of-the-art methods to perform the classification are quite imprecise. The project began with an objective to maximize the classification accuracy by modifying/fine-tuning the base model provided as part of the Hateful Memes Challenge. But, in addition to the above, we created our own multimodal models with different text and image classifiers, and extended our dataset to improve the accuracy of classification of memes into hateful or non-hateful. The model is then extended as a service where users can upload a meme and the application would inform the user if the meme is hateful or not.