Detecting Hate Speech on Memes using FixEfficientNet-L2

In Indonesia, there were 3,325 cases of hate speech spreading in 2017, based on the popularity of memes and the number of hate speech cases. There are studies that have investigated this topic and get an accuracy of 50-52%. This study will create an image classification model using the FixEfficientNet-L2 architecture with the aim of being able to identify memes that contain hate speech and memes that do not contain hate speech. hate speech. Our research had used a dataset provided by the Facebook AI team, totaling 9,540 data of meme image, and will be divided into three parts of training, validation, and data testing. Our model will use 15 different configurations on the train and test image size. Previously, preprocessing stages were carried out on the data so that the data could be studied by the model. The highest accuracy results obtained from the FixEfficientNet-L2 model with several different configurations in classifying memes are 54.8% for validation and 63% for testing, the accuracy has increased from previous studies of 2% for validation and 10% for testing.