Adapting TrOCR for Printed Tigrinya Text Recognition with Word-Aware Loss Weighting

Forum topic · 小凯 · 2026-04-24

Summary

A new arXiv paper (2604.20813) by Yonatan Haile Medhanie and Yuanhua Ni presents the first adaptation of the Transformer-based OCR model TrOCR for printed Tigrinya text written in the Ge'ez script. The authors extend the model's byte-level BPE tokenizer to cover 230 Ge'ez characters and introduce Word-Aware Loss Weighting to fix systematic word-boundary failures caused by Latin-centric BPE conventions. Without modification, TrOCR produces no usable output on Ge'ez text. After adaptation, the TrOCR-Printed variant achieves a 0.22% Character Error Rate and 97.20% exact match accuracy on a held-out test set of 5,000 synthetic images from the GLOCR dataset. Ablation studies confirm that Word-Aware Loss Weighting is the key component, reducing CER by two orders of magnitude compared to vocabulary extension alone. The full pipeline trains in under three hours on a single 8GB consumer GPU, and all code, model weights, and evaluation scripts are publicly released.

Paper Overview

Research Area: Computer Vision (CV) Authors: Yonatan Haile Medhanie, Yuanhua Ni Published: 2026-04-22 arXiv: 2604.20813

Abstract

Transformer-based OCR models have shown strong performance on Latin and CJK scripts, but their application to African syllabic writing systems remains limited. This paper presents the first adaptation of TrOCR for printed Tigrinya using the Ge'ez script. Starting from a pre-trained model, the authors 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 script.

Key Results

The unmodified model produces no usable output on Ge'ez text.
After adaptation, the TrOCR-Printed variant achieves 0.22% Character Error Rate (CER) and 97.20% exact match accuracy on a held-out test set of 5,000 synthetic images from the GLOCR dataset.
An ablation study confirms that Word-Aware Loss Weighting is the key component — it reduces CER by two orders of magnitude compared to vocabulary extension alone.
The full pipeline trains in under three hours on a single 8GB consumer GPU.
All code, model weights, and evaluation scripts are publicly released.

Source

Paper: Adapting TrOCR for Printed Tigrinya Text Recognition: Word-Aware Loss Weighting (arXiv:2604.20813)