Offensive Language Detection in Social Media Text Using N-Gram Features and Multinomial Naïve Bayes

Abstract

Offensive language on social-media and messaging platforms creates a need for moderation methods that are accurate, fast, and transparent. This study evaluates a count-based n-gram and Multinomial Naïve Bayes pipeline for binary offensive-language detection and demonstrates its integration into a Telegram bot. The experiments used the 13,240-instance training partition of the Offensive Language Identification Dataset, comprising 8,840 non-offensive and 4,400 offensive English tweets. Text cleaning, lowercasing, stop-word removal, lemmatization, and unigram-to-trigram vectorization were applied. The evaluation compared imbalance-handling strategies, 70/30, 80/20, and 90/10 train-test splits, and three n-gram configurations. The 90/10 split achieved 68.28% accuracy. The selected unigram-bigram configuration achieved 72% accuracy, with 0.57 precision, 0.65 recall, and a 0.61 F1-score for the offensive class. These results demonstrate a trade-off between overall accuracy and sensitivity to offensive messages. The Telegram component is presented as a functional proof of concept rather than a production evaluation. The study demonstrates a lightweight and interpretable prototype while acknowledging limitations related to context, dataset scope, and evolving language.