The article presents a multi-expert system for historical Manchu Optical Character Recognition (OCR) that addresses the challenge of limited labeled data and diverse visual styles, including regular script, running script, and semi-cursive chancery hand. The approach reuses checkpoints from an iterative fine-tuning process as domain specialists and employs a lightweight page-level image classifier to dispatch pages based on their visual style.
- On frozen test sets, the routed system achieved 0.30 percent Character Error Rate (CER) on regular script, 1.57 percent on memorials, and 4.83 percent on running script.
- The router attained 99.3 percent page-level domain accuracy, matching the domain-label oracle at the same precision.
- Two of the three selected specialists were not trained specifically for their final domain, demonstrating the system's ability to reuse existing checkpoints effectively.
The authors report the evaluation protocol, router design, and per-page predictions to ensure the comparison is reproducible.