Multi-Expert Routing for Low-Resource Manchu OCR

What changed

This research introduces a novel multi-expert routing system designed to address the challenges of Optical Character Recognition (OCR) for historical Manchu documents. The core problem lies in the significant visual diversity of Manchu script, which includes regular script, running script, and the semi-cursive chancery hand found in palace memorials, all within a low-resource data context. Traditional OCR methods often train a single recognizer per language or collection, which struggles when faced with such intra-language visual variations. The proposed system leverages checkpoints from an iterative fine-tuning process, treating them as domain specialists. A lightweight page-level image classifier is employed to dynamically dispatch each page to the most appropriate specialist based on its visual style. If the existing pool of checkpoints lacks a suitable expert for a particular domain, a new specialist is trained. This method aims to improve OCR accuracy by matching pages to experts specifically tuned for their visual characteristics, rather than relying on a generalist model that compromises performance across different styles.

Why it matters for builders

For AI builders working with historical documents or specialized datasets, this research presents a practical framework for tackling multi-domain OCR challenges with limited labeled data. The multi-expert routing strategy demonstrates how to effectively reuse and adapt existing OCR models (checkpoints) as specialized agents. This can significantly reduce the need for extensive retraining or the development of entirely new models for each distinct visual style within a corpus. The system's ability to dynamically select specialists at inference time offers a flexible and efficient approach to improving recognition accuracy for diverse document types.

Practical impact

The multi-expert routing system achieved notable performance on three frozen test sets of historical Manchu. The routed system matched the performance of the best specialist for each style with two-decimal precision, reporting 0.30% Character Error Rate (CER) on regular script, 1.57% CER on memorials, and 4.83% CER on running script. Crucially, the system achieved 99.3% page-level domain accuracy, effectively identifying the correct visual style for each page and routing it to the appropriate expert. This performance level surpasses what any single, generalist checkpoint could achieve. Notably, two of the three selected specialists were not originally trained for their final assigned domains, highlighting the system's adaptability. The running-script expert was the only one purpose-trained for its target domain. The researchers have made the evaluation protocol, router design, and per-page predictions publicly available to ensure reproducibility, allowing other researchers and developers to build upon this work or apply similar techniques to their own low-resource, multi-domain OCR projects.

Caveats and source limits

The research focuses specifically on historical Manchu OCR and its unique visual script variations. While the methodology is presented as generalizable, direct application to other languages or document types would require adaptation and re-evaluation. The source does not provide details on the computational overhead introduced by the page-level image classifier or the routing mechanism beyond stating it's a "lightweight" addition. Furthermore, the performance metrics are based on "frozen test sets," and the system's behavior on unseen or dynamically evolving data distributions is not detailed. The availability and cost of training additional experts, should the existing pool be insufficient, are also not specified. The research is presented as a pre-print on arXiv, and has not yet undergone formal peer review.