Over the course of four months, the researchers gathered 2,000 memes from Facebook, Twitter, and Telegram in order to create Amharic meme hate speech recognition algorithms. They use a web-based application to annotate each meme with native Amharic speakers, resulting in a Fleiss’ kappa value of 0.50. We construct unimodal and multimodal models including CNN, BiLSTM, and LSTM using various feature extraction approaches, such as word2Vec for textual information and VGG16 for pictures. With an accuracy of 63% for text and 75% for multimodal information, the BiLSTM model performs the best. The accuracy of the CNN model is 69% in image-only studies. When it comes to identifying hate speech in Amharic in memes, multimodal models perform better.

https://aclanthology.org/2024.trac-1.10

By author

Leave a Reply

Your email address will not be published. Required fields are marked *