AUTOMATED DETECTION OF OFFENSIVE TEXTS BASED ON ENSEMBLE LEARNING AND HYBRID DEEP LEARNING TECHNIQUES

Abstract

The impact of communication through social media is currently considered a significant social issue. This issue can lead to inappropriate behavior using social media, which is referred to as cyberbullying. The accessibility and freedom of expression afforded by social media platforms enable individuals to share their emotions and opinions, but it also leads to cyberbullying behavior. Automated systems are capable of efficiently identifying cyberbullying and performing sentiment analysis on social media platforms. In this dissertation, our focus is on enhancing a system to detect cyberbullying in various ways. Therefore, we apply natural language processing techniques utilizing artificial intelligence algorithms to identify offensive texts in various public datasets. The first approach leverages two deep learning models to classify a large-scale dataset, combining two techniques: data augmentation and the GloVe pre-trained word representation method to improve training performance. In addition, we utilized multi-classification algorithms on a cyberbullying dataset to identify six types of cyberbullying tweets. Our approach achieved high accuracy, particularly with TF-IDF (bigram) feature extraction, compared to previous experiments and traditional machine learning algorithms applied to the same dataset. We employed two ensemble machine learning methods with the TF-IDF feature extraction technique, which demonstrated superior classification performance. Moreover, we used four feature extraction methods, BoW, TF-IDF, Word2Vec, and GloVe, to determine which works best with the ensemble technique. Finally, we utilize a multi-channel convolutional neural network (CNN) enhanced with an attention mechanism and optimized using a focal loss function.