ENHANCING SENTIMENT ANALYSIS FOR HAUSA LANGUAGE WITH IMPROVED HAUSA TEXT STEMMER (HTS) AND MACHINE LEARNING MODELS

Abstract

This study enhances sentiment analysis for Hausa, a low-resource language spoken by over 86 million people, by introducing an improved Hausa Text Stemmer (HTS). The proposed algorithm addresses the language’s complex morphology—including prefixes, suffixes, infixes, and confixes—while also expanding common abbreviations and removing stop words. These steps improve text consistency and reduce noise, enabling more accurate feature extraction for sentiment classification. Using the AfriSenti dataset, the study evaluates four classical machine learning models—Support Vector Machine, Naïve Bayes, Gradient Boosting, and Random Forest—with performance measured by accuracy, precision, recall, and F1-score. Comparative tests against two existing stemmers demonstrate the superiority of the proposed HTS, with Gradient Boosting achieving 93.75% accuracy, significantly outperforming baseline accuracies of 86.7% and 73.1%. The findings confirm that the HTS effectively handles key linguistic challenges in Hausa, such as confixes, abbreviations, and stop words, leading to more robust sentiment classification. This work contributes valuable NLP resources for low-resource languages and underscores the importance of tailored preprocessing in sentiment analysis.