An AI-Based Adaptive Learning Platform for Multilingual and Low-Resource Educational Contexts: A Case Study on Nigeria

Synthesis: Nwogo et al. (2026) build and evaluate an AI-based adaptive learning platform for multilingual and low-resource educational contexts, with a case study on Nigerian Pidgin English. The system integrates fine-tuned large language models (LLMs) within a personalized-and-adaptive-learning (PAL) framework, using a curated Nigerian Pidgin corpus to fine-tune an instruction-tuned model. The study systematically investigates model optimization through multi-level quantization (4-bit, 5-bit, 8-bit), showing that higher-bit quantization improves semantic preservation and structural coherence while lower-bit models offer reduced inference latency with minimal degradation in instructional quality. This yields a deployable, resource-aware intelligent learning system that balances semantic robustness, cultural relevance, and computational efficiency — an experimentally validated framework for adapting LLMs to low-resource languages at practical scale.

The low-resource multilingual problem

Educational platforms in under-resourced, multilingual settings (such as Nigeria) often struggle with limited personalization, inadequate language support, and weak curriculum internationalization — reducing learner engagement and inclusivity. The paper frames this as both a linguistic and a computational challenge: learners need content in languages they actually use, but fine-tuning and serving LLMs for low-resource languages must respect severe computational constraints. This connects directly to the knowledge base's concerns about the Digital Divide, Global South equity, and Equity.

Platform design and the Nigerian Pidgin corpus

The system's core is a personalized and adaptive learning (PAL) framework that integrates fine-tuned LLMs. To achieve linguistic alignment, the authors developed a curated Nigerian Pidgin corpus and used it to fine-tune an instruction-tuned LLM, tailoring generation to the local language and context — a concrete instantiation of culturally relevant and locally grounded educational AI.

The quantization trade-off

The study's distinctive empirical contribution is a systematic analysis of multi-level quantization (4-bit, 5-bit, 8-bit) and its trade-offs:

• Higher-bit quantization (8-bit): improves semantic preservation and structural coherence — output stays closer to the full-fidelity model.
• Lower-bit models (4/5-bit): offer reduced inference latency with only minimal degradation in instructional quality — more deployable on constrained hardware.

Evaluation combined automatic semantic metrics (BLEU, ROUGE-L, BERTScore, perplexity, lexical diversity) with human-centered cultural assessment by native speakers, grounding the technical results in actual linguistic and cultural acceptability rather than metric-only scores.

What this means for practice

• Edtech designers. Choose the quantization level deliberately: 8-bit preserved semantic structure and coherence best, while the 4- and 5-bit models cut inference latency with only minimal degradation in instructional quality, so match bit width to the deployment hardware.
• Edtech designers. Validate cultural acceptability with native speakers rather than relying on automatic metrics alone — the study paired BLEU, ROUGE-L, BERTScore, perplexity, and lexical diversity with human cultural assessment by native speakers.
• Institutions. Treat the locally curated corpus as the unit of investment: fine-tuning used a 416,343-entry Nigerian Pidgin corpus scraped from BBC Pidgin and Prime9ja, showing that a low-resource language can be served without massive cloud infrastructure.
• Institutions. Plan for constrained connectivity, since offline and low-bandwidth functionality is named as necessary to reach learners in under-resourced settings.

Limitations

• Controlled setting, no field deployment: the evaluation ran in a controlled experimental setting, and large-scale longitudinal deployment in formal institutions was out of scope.
• Proxy outcomes: learner engagement and learning outcomes were assessed through short-term interactions and proxy measures rather than extended academic performance over time.
• One language only: Nigerian Pidgin English was the single representative low-resource language tested, so generalization to Nigeria's more than 520 indigenous languages is untested.
• Small evaluation set: the quantitative comparison used 14 sample prompts, which the authors describe as a baseline for comparative analysis rather than an absolute ground truth, and the fine-tuning corpus was scraped from news platforms rather than classroom material.

Citation

Nwogo, E. U., Ihianle, I. K., Machado, P., Bird, J. J., Lotfi, A., Shuaib, A. A., Akinwumi, I. I., & Oluranti, J. (2026). An AI-Based Adaptive Learning Platform for Multilingual and Low-Resource Educational Contexts: A Case Study on Nigeria. [cs.CY]. https://doi.org/10.48550/arXiv.2608.15738