Multi-Expert Routing for Multi-Domain Low-Resource OCR: A Manchu Case Study
Ranking
Overall
55
Content
60
Popularity
44
Observed public metrics from 1 member.
Merged summary
TL;DR — A study on historical Manchu OCR that routes document pages to style-specialized recognition experts via a lightweight page-level image classifier, showing that reusing fine-tuning checkpoints as domain specialists can match dedicated per-style models even with limited labeled data.
- Handles three visually distinct writing styles (regular script, semi-cursive chancery/palace memorials, running script) by dispatching pages with an image classifier to the best expert, training a new expert only when the checkpoint pool lacks a suitable one.
- The routed system matches the selected per-style specialist to two-decimal precision: 0.30% CER (regular), 1.57% (memorials), 4.83% (running script).
- The router hits 99.3% page-level domain accuracy, matching a domain-label oracle at the same precision.
- Notably, 2 of 3 chosen specialists were not originally trained for their final domain—checkpoints from iterative fine-tuning generalized as reusable experts; authors release protocol, router design, and per-page predictions for reproducibility.
Sources (1)
Multi-Expert Routing for Multi-Domain Low-Resource OCR: A Manchu Case Study
Public signals
Semantic Scholar citations 0 · Semantic Scholar influential citations 0
TL;DR — A study on historical Manchu OCR that routes document pages to style-specialized recognition experts via a lightweight page-level image classifier, showing that reusing fine-tuning checkpoints as domain specialists can match dedicated per-style models even with limited labeled data.
- Handles three visually distinct writing styles (regular script, semi-cursive chancery/palace memorials, running script) by dispatching pages with an image classifier to the best expert, training a new expert only when the checkpoint pool lacks a suitable one.
- The routed system matches the selected per-style specialist to two-decimal precision: 0.30% CER (regular), 1.57% (memorials), 4.83% (running script).
- The router hits 99.3% page-level domain accuracy, matching a domain-label oracle at the same precision.
- Notably, 2 of 3 chosen specialists were not originally trained for their final domain—checkpoints from iterative fine-tuning generalized as reusable experts; authors release protocol, router design, and per-page predictions for reproducibility.