Machine Learning Enhanced By Sentiment Analysis For Cyberbullying Detection Using Nlp And Lstm

Abstract

The proliferation of social media has brought the critical challenge of cyberbullying to the forefront of digital safety, causing significant psychological harm. Traditional machine learning detection methods often fail to capture the nuanced and contextual nature of harmful language. This paper proposes an advanced framework that combines Natural Language Processing (NLP) techniques with a Long Short-Term Memory (LSTM) network for improved cyberbullying detection in online text. The methodology applies refined text preprocessing steps—tokenization, stop word removal, stemming, and lemmatization—to ensure high-quality input data. Sentiment features and contextual patterns are then extracted using word embeddings to preserve semantic information. These processed inputs are fed into an LSTM model, which effectively captures sequential and temporal dependencies, making it well-suited for understanding the dynamic language of cyberbullying. Resampling techniques are also employed to address multi-class data imbalance, enhancing model robustness. The proposed system demonstrates that integrating deep learning with comprehensive NLP significantly enhances accuracy and contextual understanding, paving the way for more effective and reliable automated content moderation. Experimental results show the proposed LSTM model achieves 92% accuracy, outperforming conventional classifiers such as Naive Bayes and Extra Trees.