Transformer-based coreference resolution modeling for Amharic text

Abstract

Abstract Coreference resolution is the task of identifying all textual expressions that refer to the same real-world entity within a document. Coreference resolution is essential for several NLP applications, including machine translation, question answering, information extraction, summarization, and dialogue systems. While substantial progress has been made in high-resource languages such as English and Arabic. Amharic is a morphologically rich, low-resource language remains underexplored due to the scarcity of annotated datasets and specialized language tools. This study proposes a transformer-based approach for Amharic coreference resolution using multilingual BERT (mBERT ) . The proposed system combines contextual embeddings extracted from mBERT with manually engineered mention-pair features, such as sentence-level distance and exact string matching, through a supervised mention-pair classification framework. The system is composed of two main phases: training and testing. The training phase includes preprocessing, tokenization, sentence segmentation, feature extraction, and model training. The testing phase involves similar preprocessing steps and the prediction of coreference clusters. We created a novel annotated Amharic corpus consisting of 312 documents, 5428 sentences, 18,763 mentions, and 4912 coreference clusters to train and evaluate the model. Standard coreference evaluation metrics, including MUC, B 3 , CEAF-m, CEAF-e, and BLANC, were employed. The results show that the proposed model outperforms existing state-of-the-art Amharic systems, achieving F1-scores of 80%, 85.71%, 90.9%, 88.86%, and 81.7% across the respective metrics. This study demonstrates that transformer-based contextual embeddings can effectively capture the semantic relationships necessary for Amharic coreference resolution.