A GRAPH CONVOLUTIONAL NETWORK AND RECURRENT NEURAL NETWORK ENSEMBLE FOR EXTRACTIVE TEXT SUMMARISATION IN THE HAUSA LANGUAGE

Abstract

Automatic Text Summarisation (ATS) is crucial for managing information overload, especially in low-resource languages like Hausa. This study proposes a hybrid extractive approach that combines Graph Convolutional Networks (GCN) and Recurrent Neural Networks (RNN) to improve sentence ranking accuracy. By integrating GCN’s structural learning with RNN’s sequential modeling, the method overcomes limitations of existing graph-based techniques. The research was conducted in Visual Studio IDE using Python 3.12.4, with key libraries like NLTK, Pandas, and NetworkX. A pre-processed Hausa news dataset was tokenised, normalised, and vectorised using TF-IDF and a pre-trained Hausa FastText model to build a sentence similarity graph. GCNs propagated sentence embeddings, while RNNs refined rankings by capturing sequential dependencies. Experiments on 113 Hausa news articles showed the GCN-RNN model outperformed Modified PageRank, achieving higher ROUGE-1 precision (90.00) and balanced F1-scores. The Wilcoxon Signed-Rank Test confirmed significant improvements. Despite added computational overhead, the approach remains feasible for moderate datasets, with scalability as a key future focus. This study offers a robust and contextually coherent approach to Hausa text summarisation, advancing extractive summarisation techniques and multilingual ATS research. Future work will focus on optimising model efficiency and scalability while exploring transformer-based architectures for further enhancements.