Klasifikasi Cyberbullying dengan Menggunakan Convolutional Neural Network (CNN) dan ELMo Embeddings

Abstract

The rapid development of social media in the digital era has led to various negative impacts. One of them is cyberbullying, which can cause serious psychological effects on victims. This study aims to analyze the best combination of hyperparameter tuning for a cyberbullying classification model using a Convolutional Neural Network (CNN) with Embedding from Language Model (ELMo) word representation, as well as to examine the effect of using stemming in the preprocessing stage on accuracy. The data used consists of English-language cyberbullying data obtained from the Kaggle website, with five classification categories: not_cyberbullying, gender, religion, age, and ethnicity. The results show that hyperparameter tuning produces the same best configuration for both conditions—without stemming and with stemming—namely 256 filters and a learning rate of 0.0001. The testing accuracy achieved is 0.8957 for the model without stemming and 0.9017 for the model with stemming. The testing loss is 0.2891 for the model without stemming and 0.2923 for the model with stemming. The application of stemming does not provide a significant improvement in accuracy and even increases the model loss, considering that ELMo as a contextual embedding is already capable of capturing morphological variations of words contextually, making the contribution of stemming limited.