Search papers, labs, and topics across Lattice.
This study introduces a multi-expert system for low-resource Optical Character Recognition (OCR) of historical Manchu texts, which leverages domain-specific checkpoints and a lightweight image classifier to effectively route pages based on their visual styles. The approach demonstrates remarkable precision in character error rates (CER) across various scripts, achieving 0.30% for regular script, 1.57% for memorials, and 4.83% for running script, while maintaining a 99.3% accuracy in page-level domain classification. Notably, two of the three specialists performed effectively without being specifically trained for their final domains, highlighting the system's adaptability and efficiency in resource-constrained settings.
Achieving 99.3% accuracy in domain classification, this multi-expert OCR system adapts to diverse Manchu writing styles with minimal training data.
Historical Manchu OCR must accommodate various visually distinct writing styles, including regular script, running script, and the semi-cursive chancery hand used in palace memorials, despite limited labeled data. We study a multi-expert system that reuses checkpoints from an iterative fine-tuning process as domain specialists and uses a lightweight page-level image classifier to dispatch pages by visual style. When the checkpoint pool lacks a suitable specialist, we train an additional expert for that domain. On three frozen test sets, the routed system matches the selected specialist for each style at two-decimal precision: 0.30 percent CER on regular script, 1.57 percent on memorials, and 4.83 percent on running script. The router achieves 99.3 percent page-level domain accuracy and matches the domain-label oracle at the same precision. Two of the three selected specialists were not trained specifically for their final domain; only the running-script expert was trained with that domain as its target. We report the evaluation protocol, router design, and per-page predictions to make the comparison reproducible.
