Multi-expert routing adapts low-resource Manchu OCR to historical scripts

Introduction

Historical OCR is rarely a single-domain recognition problem. A collection may use the same writing system while varying widely in visual form, layout, and calligraphic convention. Manchu historical documents illustrate this challenge well: regular script, running script, and the semi-cursive chancery hand seen in palace memorials can look different enough that a model trained for one style may not be the best fit for another.

The arXiv paper “Multi-Expert Routing for Multi-Domain Low-Resource OCR: A Manchu Case Study” addresses this setting with a practical multi-expert design. Instead of assuming that one OCR model should handle every page equally well, the system first classifies the page’s visual domain and then sends it to an appropriate OCR specialist.

Key points

• A low-resource, multi-domain OCR problem: The study focuses on historical Manchu OCR, where labeled data is limited and the target pages span visually distinct styles including regular script, running script, and palace memorial handwriting.
• Experts from existing checkpoints: The system reuses checkpoints produced during an iterative fine-tuning process as candidate domain specialists. This turns intermediate training artifacts into useful models rather than discarding them.
• A lightweight page-level router: Before OCR, a compact image classifier predicts the page’s domain based on visual style. The paper reports 99.3% page-level domain accuracy for this router.
• Specialist-level performance after routing: On three frozen test sets, the routed system matches the selected specialist for each style at two-decimal precision: 0.30% CER on regular script, 1.57% on memorials, and 4.83% on running script. It also matches the domain-label oracle at the same precision.
• Not every expert was trained for its final role: Two of the three selected specialists were not originally trained specifically for the final domain they served. Only the running-script expert was trained with that domain as its target, suggesting that checkpoints can acquire transferable strengths along the fine-tuning path.

Why it matters

The paper’s main contribution is not simply another OCR score, but a workflow for dealing with fragmented, scarce, and stylistically diverse historical data. Many cultural heritage collections face exactly this situation: documents come from different periods, hands, and institutional contexts, while ground-truth transcription remains expensive.

A routed multi-expert approach offers a middle ground between a single universal model and a fully separate model for every niche domain. If the router is accurate and the expert pool is well curated, the system can exploit specialization without requiring large labeled datasets for every style.

The authors also emphasize reproducibility by reporting the evaluation protocol, router design, and per-page predictions. The results are specific to the Manchu case study, so broader claims across other scripts or noisier archives still require further testing. Even so, the design principle is broadly relevant: in low-resource OCR, identifying what kind of page you are reading may be as important as the recognizer itself.

Source: arXiv