Amharic Hate‐Speech Detection and Classification: A PRISMA 2020 Systematic Review

The proliferation of social media has accelerated the spread of hate speech, posing significant societal and computational challenges, particularly for low‐resource languages such as Amharic, a morphologically rich Ethio‐Semitic language written in the Fidel script. This study presents a PRISMA‐2020 systematic literature review of computational approaches to Amharic hate‐speech detection and classification, with the aim of synthesizing empirical evidence, identifying methodological trends, and highlighting research gaps. A systematic search was conducted across seven electronic databases (ACL Anthology, ACM Digital Library, Scopus, Web of Science, IEEE Xplore, ScienceDirect, and SpringerLink) supplemented by a grey‐literature search, covering records published between January 1, 2016, and June 30, 2026. Eligibility was defined using the PICOC framework, and study selection, data extraction, and quality assessment were performed independently by two reviewers, yielding inter‐rater reliability of Cohen’s κ  = 0.82 at title/abstract screening and κ  = 0.77 at full‐text assessment. Methodological quality was evaluated through a four‐dimension framework (dataset transparency, methodological rigor, evaluation completeness, and reproducibility), and the certainty of evidence was assessed using GRADE‐CERQual. A total of 36 studies met the inclusion criteria. The findings reveal a clear progression from traditional machine learning methods to recurrent neural network architectures and, more recently, to transformer‐based and large‐language‐model approaches. Quality assessment showed that 53% of studies ( n  = 19) achieved higher methodological quality, 42% ( n  = 15) medium, and 6% ( n  = 2) low, with reproducibility emerging as the weakest dimension (mean 1.78 of 3) and only 19% of studies releasing data or models publicly. Although several studies report high accuracy (up to 98%), these results are often inflated by small, domain‐specific datasets, class imbalance, and inconsistent evaluation practices; on curated corpora, reported F1 values in the approximately 70%–81% range occur among several of the better‐documented curated benchmarks reviewed here; this range should not be interpreted as a universal performance ceiling. The review further identifies pervasive challenges related to reproducibility, transparency, annotation variability, and limited consideration of fairness, bias, and dual‐use risks. Overall, while recent advances, particularly transformer‐based and large‐language‐model methods, demonstrate promising technical progress, the field remains constrained by data scarcity and methodological inconsistencies. Future research should prioritize the development of larger, publicly available, and ethically governed datasets, the adoption of robust and context‐aware modeling approaches, and the systematic integration of fairness‐aware evaluation frameworks.