Novel Approach to Offensive Language Detection on Social Media: Tree CNN Integration with Adversarial Bi-LSTM

Offensive language detection on social media platforms is a crucial task for maintaining a healthy online environment
and ensuring user safety. Traditional methods often struggle to effectively capture the nuances and dynamic nature of offensive language in such diverse and rapidly evolving contexts. Our research presents a novel approach to offensive language detection on social media by integrating Tree Convolutional Neural Networks (CNN) with Adversarial Bidirectional Long Short-Term Memory (Bi-LSTM) networks. We address the challenges of imbalanced data and semantic understanding by employing the Synthetic Minority Over-sampling Technique (SMOTE) and Word2Vec for feature extraction. To enhance model interpretability and focus on relevant features, we incorporate attention mechanisms within both the Tree CNN and the Adversarial Bi-LSTM.
We utilize two attention mechanisms: one for identifying repetitive patterns using an Entropy Pruning Method and another for error loss monitoring during training. This dual attention mechanism enables our model to effectively distinguish offensive from non-offensive text while also providing insights into model decisions. Experimental findings on benchmarking social media datasets show that our suggested methodology outperforms cutting-edge approaches in terms of accuracy and interpretation. Our study contributes to advancing offensive language detection techniques on social media platforms and provides a framework for developing more reliable and interpretable models for online content moderation.