Resource-Efficient Somali News Classification: A Supervised Linear SVM Approach on a Novel Text Corpus

Abstract The rise of media means more people are creating content online than ever before. This has caused a surge in the amount of content, on the Internet. There is still a significant imbalance in automated organizing tools in favor of high-resource languages. Low-resource languages such as Somali suffer from serious technological disenfranchisement due to an extreme lack of data sets and unique linguistic features, including complex morphology and orthographic variation. In order to bridge this digital divide, this paper describes the development of a supervised machine learning algorithm based on a novel manually curated benchmark data set consisting of 7,241 Somali news stories. Although unsupervised algorithms such as Latent Dirichlet Allocation (LDA) and Latent Semantic Indexing (LSI) produced structurally poor classifications, a supervised Support Vector Machine (SVM) pipeline with tailored tokenization, stopword elimination, and Term Frequency-Inverse Document Frequency (TF-IDF) feature selection resulted in outstanding results. The classifier developed in this study was able to accurately classify texts into eight primary thematic categories (Caafimaad, Ciyaaro, Dagaal, Diini, Ganacsi, Madadaalo, Siyaasad, Tiknoolajiyad). with an accuracy of 95%. These results show that computationally efficient traditional machine learning algorithms can be a mathematically sound and resource-friendly alternative to computationally heavy deep learning systems in low-resource settings.