Advancing Amharic Information Retrieval: A Comparative Analysis of Traditional, Neural, and Transformer-Based Models

Amharic, the official language of Ethiopia, presents unique challenges for information retrieval (IR) due to its complex morphology, vocabulary richness, and limited computational and linguistic resources. This study evaluates and compares traditional retrieval methods (BM25, TF-IDF), shallow neural models (FastText, word2vec), and transformer-based dense retrieval models (SPLADE) to address these challenges. We explore query expansion techniques using WordNet and pre-trained embeddings, highlighting the impact of morphological analysis and learned text representations on retrieval effectiveness. Experiments on two Amharic datasets demonstrate the superiority of semantics-based retrieval over traditional methods, although transformer-based should limited impact. This work provides actionable insights for advancing semantic IR in low-resource languages like Amharic.