Parameter-Efficient Fine-Tuning of the Google Gemma-7B Model for Hausa Language Sentiment Analysis Task Using Low-Rank Adaptation Technique

For low-resource languages like Hausa, the lack of annotated data is a major obstacle to developing reliable sentiment analysis models. Conventional approaches rely on fine-tuning big language models, which are inappropriate for such languages since they demand large amounts of data and processing power. The Google Gemma-7b model is optimised for Hausa sentiment analysis in this study by employing LowRank Adaptation (LoRA), a parameter-efficient fine-tuning technique. By combining low-rank matrices and freezing the pre-trained model weights, LoRA minimises the number of trainable parameters, allowing for efficient adaptation while lowering computing expenses. Accuracy rose from 37.7 % to 66.0 %, and the refined model, which was trained on the NaijaSenti dataset, showed notable improvements in F1-scores across sentiment classes, especially for positive emotions (F1-score: 0.77). This work advances Natural Language Processing (NLP) for under-represented languages by providing a scalable and effective way to adapt big language models to jobs with limited resources. This work addresses language inclusivity and computational efficiency, laying the groundwork for future studies in low-resource NLP.