Resource-efficient somali news classification using TF-IDF and linear SVM: a supervised baseline on a curated news corpus

Abstract The rapid growth of digital journalism has increased the need for automated systems that can organize large volumes of textual content. However, mature natural language processing resources remain concentrated in high-resource languages, while low-resource languages such as Somali continue to face limited availability of annotated datasets, language-specific preprocessing tools, and reproducible benchmarks. This study develops a lightweight supervised machine learning pipeline for Somali news topic classification using a manually curated corpus of 7240 articles collected from BBC Somali and VOA Somali. The articles were organized into eight mutually exclusive topic categories: Caafimaad, Ciyaaro, Dagaal, Diini, Ganacsi, Madadaalo, Siyaasad, and Tiknoolajiyad. The final model used word-level TF-IDF unigram features with a Linear Support Vector Machine classifier. In a stratified 80/20 holdout evaluation, the best Linear SVM configuration achieved 94.20% accuracy. Additional repeated stratified holdout testing and stratified five-fold cross-validation produced mean accuracies of 93.11% and 93.20%, respectively, indicating stable performance across data partitions. Supervised baseline comparisons showed that Linear SVM outperformed Multinomial Naive Bayes, Logistic Regression, Random Forest, and alternative SVM n-gram variants under the same preprocessing and evaluation conditions. Computational profiling further showed that the final pipeline trained in under two seconds, classified the test set in less than one quarter of a second, produced a serialized model below 1 MB, and required no GPU acceleration. The findings position TF-IDF + Linear SVM as a strong, transparent, and resource-efficient baseline for Somali news classification, while future work should compare this baseline with neural and transformer-based models under the same data, preprocessing, and evaluation protocol.