A specific architectural tweak turned a completely non-functional AI model into one that reads the Tigrinya language with 97.2% accuracy.
April 24, 2026
Original Paper
Adapting TrOCR for Printed Tigrinya Text Recognition: Word-Aware Loss Weighting for Cross-Script Transfer Learning
arXiv · 2604.20813
The Takeaway
Most global AI models produce zero usable output for underserved languages like Tigrinya, which uses the Ge'ez script. By implementing a word-aware loss weighting system, researchers were able to successfully bridge this digital gap. This approach focuses on how the model learns individual words rather than just random characters. It proves that we do not need massive new datasets to include the whole world in the AI revolution. A simple change in how the model is trained can unlock access for millions of people. Language accessibility is now an engineering problem with a clear solution.
From the abstract
Transformer-based OCR models have shown strong performance on Latin and CJK scripts, but their application to African syllabic writing systems remains limited. We present the first adaptation of TrOCR for printed Tigrinya using the Ge'ez script. Starting from a pre-trained model, we extend the byte-level BPE tokenizer to cover 230 Ge'ez characters and introduce Word-Aware Loss Weighting to resolve systematic word-boundary failures that arise when applying Latin-centric BPE conventions to a new s