Multi-Expert Routing for Multi-Domain Low-Resource OCR: A Manchu Case Study
2026-07-15 • Computer Vision and Pattern Recognition
Computer Vision and Pattern RecognitionArtificial IntelligenceMachine Learning
AI summaryⓘ
The authors studied how to improve optical character recognition (OCR) for historical Manchu texts, which come in different handwriting styles. They created a system that uses multiple specialized models, each fine-tuned for a different style, and a classifier that decides which model to use for each page. Their system performs very well, nearly matching models trained specifically for each style, even when some experts weren't originally trained on that style. They provide detailed results and methods to allow others to replicate their work.
Manchu scriptOCRregular scriptrunning scriptsemi-cursive chancery handiterative fine-tuningmulti-expert systempage-level image classificationcharacter error ratedomain adaptation
Authors
Zhan Chen, Jiqiao Ma, Chih-wen Kuo
Abstract
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.