Multi-Expert Routing for Low-Resource Manchu OCR

What changed

This research introduces a multi-expert routing system for Optical Character Recognition (OCR) specifically applied to historical Manchu documents. The core challenge addressed is the need to accurately process Manchu text that exhibits significant visual variations across different writing styles, such as regular script, running script, and the semi-cursive chancery hand found in palace memorials. This is particularly difficult due to the scarcity of labeled data for training robust OCR models.

The proposed system employs a multi-expert architecture. It reuses existing OCR model checkpoints, which have undergone iterative fine-tuning, as specialized domain experts. A lightweight page-level image classifier is then used to analyze incoming pages and dispatch them to the most appropriate specialist based on their visual style. If the existing pool of specialists does not contain a model suitable for a particular domain, a new expert is trained for that specific style. This method aims to efficiently adapt OCR capabilities to diverse scripts without requiring extensive new data for every variation.

The system's performance was evaluated on three distinct, frozen test sets. The results indicate that the routed system achieves high accuracy in matching pages to the correct specialist. Specifically, it reports character error rates (CER) of 0.30% on regular script, 1.57% on memorials, and 4.83% on running script. Furthermore, the routing mechanism itself demonstrates strong performance, achieving 99.3% accuracy at the page level in domain classification, aligning with an oracle that already knows the correct domain label. Notably, two of the three specialists used in the final configuration were not initially trained with their assigned domain as their primary target, highlighting the system's adaptability and the effectiveness of reusing checkpoints.

To ensure reproducibility, the researchers have detailed the evaluation protocol, the design of the router, and provided per-page predictions.

Why it matters for builders

This work presents a valuable blueprint for building robust OCR systems in low-resource scenarios, especially for historical or specialized document types. The multi-expert routing strategy, combined with the reuse of fine-tuned checkpoints, offers an efficient way to handle diverse visual inputs without the prohibitive cost of training separate, large models for each variation. Builders working on document digitization, archival projects, or any application requiring OCR for scripts with limited digital resources can adapt this methodology.

Practical impact

The practical impact of this research lies in its potential to unlock access to historical archives and documents that are currently difficult to digitize and analyze due to script variations and data limitations. For Manchu language studies, this could mean a significant increase in the volume of texts that can be processed and searched. Beyond Manchu, the principles of multi-expert routing and adaptive specialist selection can be applied to other low-resource OCR challenges, such as historical manuscripts in various languages, specialized scientific notations, or even unique handwritten forms. The reported accuracy metrics suggest a viable path towards practical deployment in archival and research settings.

Caveats and source limits

The primary source for this information is a research paper published on arXiv. The findings are based on a specific case study involving Manchu OCR and three distinct writing styles. While the methodology is presented as generalizable, its effectiveness on other languages or script types would require further validation. The research focuses on the routing and specialist selection mechanism, and the performance is reported in terms of character error rates and domain accuracy. The paper does not provide details on the computational resources required for training or inference, nor does it offer a direct comparison with end-to-end models trained on massive, diverse datasets (which may not be feasible for low-resource scenarios). The date of publication (2026) indicates this is a forward-looking research output.