AN IMPROVED MULTINOMIAL NAÏVE BAYES ALGORITHM WITH LAPLACE SMOOTHING FOR NEWS CLASSIFICATION: THE BISALLAH NEWS NETWORK MODEL (BNNM)

Abstract

Naïve Bayes remains one of the most widely used algorithms for text classification and sentiment analysis, yet its application to Nigerian newspaper social-media data has not been previously documented. This study develops and evaluates an improved Multinomial Naïve Bayes classifier - incorporating TF-IDF feature weighting and Laplace (additive) smoothing - for the automatic classification of Nigerian newspaper tweets into politics, sports, and business categories. Drawing on a cleaned corpus of 24,679 tweets from The Punch, Vanguard, and The Guardian (2015–2018), the classifier was trained on a 70/30 train-test split and tuned via grid-search cross-validation across n-gram ranges, TF-IDF normalisation schemes, and smoothing parameters (α). The resulting model - termed the Bisallah News Network Model (BNNM) - achieved an overall classification accuracy of 83.1%, with category-level prediction accuracies of 80.6% for politics, 72.1% for sports, and 96.0% for business, an average recall of 0.75, and an average F1-score of 0.77. These results are broadly comparable to, though in some cases below, prior studies (e.g., Preety & Dahiya, 2015, at 89.01%; Perdana & Pinandito, 2018, at 83.8%), with differences attributable to corpus size and composition. The study demonstrates that an improved, Laplace-smoothed Multinomial Naïve Bayes pipeline can deliver practically useful classification performance on noisy, real-world Nigerian social-media text, offering a foundation for automated content categorisation and editorial decision-support tools in African digital newsrooms.