A machine learning-based approach for information extraction and its implications for sustainable digital resilience

Recent advances in Natural Language Processing (NLP) and Machine Learning (ML) have revolutionized Information Extraction (IE) in high-resource languages, yet low-resource languages spoken by over million people in southern Ethiopia remain digitally marginalized. To the best of our knowledge, this study presents one of the first annotated corpora and benchmark experiments for Gamoththo information extraction. Using a manually annotated corpus of 600 sentences collected from textbooks, media, and academic sources, three supervised learning algorithms Bidirectional Long Short-Term Memory with Conditional Random Fields (Bi-LSTM + CRF), Support Vector Machine (SVM), and fine-tuned Bidirectional Encoder Representations from Transformers (BERT) were trained and tested. Results revealed that fine-tuned BERT achieved the highest NER performance (Precision: 92.3%, Recall: 90.8%, F1-score: 91.5%), followed by Bi-LSTM + CRF (F1-score: 88.9%) and SVM (F1-score: 76.8%). For Relation extraction, SVM attained superior accuracy (79.4%, F1-score: 86.4%), while Bi-LSTM + CRF excelled in event extraction (Accuracy: 80.9%). The fine-tuned BERT model achieved higher F1-scores compared to Bi-LSTM + CRF and SVM for NER, suggesting that transfer learning provides measurable improvements on this dataset. These initial benchmarks provide a foundation for future development of Gamoththo NLP tools and contribute toward broader digital inclusion efforts.