Beyond Pretraining Bias: Evaluating Language-Adaptive Fine-Tuning Strategies for Sentiment and Topic Classification in Three Nigerian Languages

This paper presents a controlled empirical study comparing four adaptation strategies — zero-shot prompting, standard fine-tuning, language-adaptive fine-tuning (LAFT), and ensemble decoding — applied across sentiment classification and news-topic classification in Hausa, Yorùbá, and Igbo. Using AfriSenti, NaijaSenti, and MasakhaNEWS benchmarks with AfriBERTa-large and mBERT as base models, we find LAFT with curated monolingual pretraining data produces consistent F1 gains of 4.2–9.8 percentage points over standard fine-tuning. We document failure modes hidden by aggregate metrics: tonal morphology confusion in Yorùbá, dialectal token misalignment in Hausa, and orthographic inconsistency in Igbo.