Comparative Analysis of Deep Learning and Transformer Models for Low Resource Hausa Language Text Summarization

Abstract

Natural language processing for low-resource languages presents unique and multifaceted challenges that necessitate carefully designed research methodologies and model architectures capable of delivering meaningful performance despite limited data availability. This paper presents a systematic comparative study of three distinct model architectures applied to the task of automatic text summarization for the Hausa language: a Recurrent Neural Network (RNN) baseline, a Long Short-Term Memory (LSTM) network, and the state-of-the-art mT5_multilingual_XLSum transformer model. Using the GEM/xlsum Hausa dataset, comprising 8,020 Hausa language news article-summary pairs, this study implements a comprehensive research pipeline encompassing dataset collection, data preprocessing with T5Tokenizer and AutoTokenizer, synonym-based data augmentation, model training with a 70:10:20 train-validation-test split, and rigorous performance evaluation using ROUGE metrics. Results demonstrate that the mT5 transformer model substantially outperforms both deep learning baselines across all evaluation metrics, with ROUGE-1, ROUGE-2, ROUGE-L, and ROUGE-LSUM scores of 63.73%, 41.54%, 51.79%, and 48.81%, respectively compared to the LSTM's 43.73%, 21.54%, 41.79%, 48.81% and the RNN's 51.73%, 26.54%, 37.79%, 38.81%. Analysis of these differential outcomes reveals architectural factors driving the performance gap, with implications for future research in low-resource NLP, multilingual model development, and computational linguistic resource creation for underrepresented African languages.