Jumia-senti: hybrid distillation and quantization framework for efficient african e-commerce sentiment analysis

Abstract

The proliferation of e-commerce in Africa has generated large amounts of user-generated text content that need to be analyze, but automated sentiment analysis in this domain remains stifled by a critical resource-utility gap. While large language models (LLMs) like AfroBERT, AfroXLM-R, and others have achieved state-of-the-art performance in text classification, their massive model size or footprint and higher inference time make them unsuitable for deployment on the resource-constrained devices prevalent in the Nigerian market. To bridge this gap, this study propose a resource-efficient framework that compresses the AfroXLM-R model into a lightweight architecture using Knowledge Distillation (KD) and static Post-Training Quantization (PTQ). We rigorously evaluate this approach on the novel Jumia product review dataset since existing datasets fail to capture the complex, code-switched vernacular ("Naija") characteristic of commercial product reviews. We evaluate our approach against several African-centric baselines.  Experimental results demonstrate that our proposed approaches significantly reduce computational overhead and we achieve an accuracy of 84.25% F1-score of 84.29 %, retaining 93% of the large AfroXLM-R model performance while reducing the model size by 33.53x (from 2260MB to 67.4MB) and accelerating inference latency by speedup 8.76 x (from 6973.49 ms to 795.98 ms). This work offers a practical blueprint for deploying culturally competent, low-latency AI solutions in infrastructure-limited environments.