Paper deep dive

Multi-Expert Routing for Multi-Domain Low-Resource OCR: A Manchu Case Study

Zhan Chen, Jiqiao Ma, Chih-wen Kuo

Year: 2026Venue: arXiv preprintArea: cs.CVType: PreprintEmbeddings: 46

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%

Last extracted: 7/18/2026, 10:35:08 AM

Summary

This paper presents a multi-expert routing system for Optical Character Recognition (OCR) of historical Manchu documents, addressing the challenge of low-resource, multi-domain data. The system utilizes a lightweight ResNet-18 image classifier as a domain router to dispatch page images to specialized Nougat model checkpoints (experts) trained on different visual styles: regular script, memorial script, and running script. The approach achieves high accuracy (99.3% domain classification) and low Character Error Rate (CER) by selecting the best-performing expert for each visual style, outperforming single-model approaches.

Entities (10)

Manchu OCR · task · 98%Character Error Rate · metric · 95%Nougat · model · 95%ResNet-18 · model · 95%Memorial Script · domain · 92%Regular Script · domain · 92%Running Script · domain · 92%v3.0 · model-checkpoint · 90%v3.0 s2 · model-checkpoint · 90%v5.0 · model-checkpoint · 90%

Relation Signals (9)

Multi-Expert Routing System → achievesaccuracy → 99.3%

confidence 95% · The router achieves 99.3 percent page-level domain accuracy

Multi-Expert Routing System → achievescer → 4.83%

confidence 95% · 4.83 percent on running script

Multi-Expert Routing System → achievescer → 1.57%

confidence 95% · 1.57 percent on memorials

Multi-Expert Routing System → achievescer → 0.30%

confidence 95% · 0.30 percent CER on regular script

Multi-Expert Routing System → routesto → Nougat

confidence 95% · All experts share one architecture—a Nougat document encoder–decoder

Multi-Expert Routing System → uses → ResNet-18

confidence 95% · The router is a small residual image classifier [11] (ResNet-18...)

v5.0 → specializesin → Running Script

confidence 92% · v5.0 for running script

v3.0 s2 → specializesin → Memorial Script

confidence 92% · v3.0 s2 for memorials

v3.0 → specializesin → Regular Script

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

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.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

44,315 characters extracted from source content.

Expand or collapse full text

Multi-Expert Routing for Multi-Domain Low-Resource OCR: A Manchu Case Study Zhan Chen 1,2 * Jiqiao Ma 3† Chih-wen Kuo 4‡ 1 Institute of Advanced Studies, Beijing Normal University, Zhuhai, China 2 Faculty of Humanities and Social Sciences, Beijing Normal University–Hong Kong Baptist University United International College (UIC), Zhuhai, China 3 Independent Scholar 4 Department of Applied History, National Chiayi University, Chiayi, Taiwan July 2026 Abstract Historical Manchu OCR must accommodate var- ious visually distinct writing styles—like regu- lar script, running script, and the semi-cursive chancery hand used in palace memorials—despite limited labeled data. We study a multi-expert sys- tem 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% CER on regular script, 1.57% on memorials, and 4.83% on running script. The router achieves 99.3% page-level do- main accuracy and matches the domain-label or- acle at the same precision. Two of the three se- lected 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 re- producible. 1 Introduction Optical character recognition is well developed for many modern, high-resource scripts, but remains difficult for languages represented mainly in his- * chenzhan4321@gmail.com † jackiemorin@163.com ‡ Corresponding author: ziwenkuo@gmail.com torical documents and manuscripts. Existing work in this setting commonly trains one recognizer per language, collection, or hand—from Tesseract [36] for print to Transkribus [16] and kraken [19] for historical manuscripts—using the limited anno- tated data available. This approach is reasonable when the visual material is homogeneous. It be- comes less suitable when a single language is rep- resented by several scripts, hands, periods, or me- dia, because the recognizer must then accommo- date substantially different image distributions. Manchu provides a clear example of this prob- lem. It was the administrative language of the Qing empire (1636–1912), and its records are impor- tant primary sources for the history of China, In- ner Asia, and the early modern world [6, 9]. The surviving material is extensive: the First Historical Archives of China holds on the order of two million documents written wholly or partly in Manchu, while the Qing archive retrieval system of the Na- tional Palace Museum in Taipei catalogues 17,056 Manchu-language records, including 2,801 palace memorials. Only a small fraction of this material has been transcribed into machine-readable text. Page-level OCR therefore offers a practical route from digitized images to searchable text, provided that it can accommodate the visual variation within the corpus. Historical Manchu is also a demanding OCR tar- get (Section 2 details the script and its material): a vertical, connected script with position-dependent letterforms, degraded archival scans, and material that spans several visually distinct domains, from the upright letterforms of woodblock print to the semi-cursive chancery hand of palace memorials 1 arXiv:2607.14041v1 [cs.CV] 15 Jul 2026 to fluent running-script manuscript. Performance on regular-script pages therefore need not transfer to running-hand cursive. This within-language variation is the central problem considered here and is common to other historical scripts. Pooling all domains into one training set can produce a useful generalist, but it also creates a compromise: in our experiments, a checkpoint that performs well on regular script performs poorly on running script, and the con- verse also occurs. The limited size of each domain makes it undesirable to train every specialist inde- pendently from scratch. Our starting point is the checkpoint history pro- duced by iterative fine-tuning. The project con- tains a sequence of versions (v3.0, v4.0, v4.1, and later versions), trained on different data snapshots and augmentation recipes. Their domain scores differ: one checkpoint performs best on regular script, another on memorials, and another on run- ning script. Changes that improve a handwriting domain can therefore reduce performance on a do- main that was already well represented. The result- ing checkpoints form an observed trade-off in the three-domain error space (Section 5). We retain these checkpoints and select com- plementary specialists from them. A lightweight image-level domain router predicts the visual style of each page and dispatches the page to the cor- responding specialist.In this finite checkpoint pool, the domain-label oracle combines the lowest observed domain-specific errors, and the learned router matches it at the reported precision. The additional computation is one classifier pass be- fore page transcription; the page recognizers them- selves are existing checkpoints. On the Manchu test sets used here, the routed system reaches sub-percent CER on regular-script pages while also evaluating memorial and running- script material. The broader implication is method- ological: in a low-resource, multi-domain project, earlier checkpoints can remain useful when their domain-specific strengths are preserved and se- lected at inference time. We report the model as- signments, router protocol, evaluation procedure, and per-page predictions so that the result can be checked and extended. Our contributions are: 1. a benchmark-scoped trade-off formulation of multi-domain low-resource OCR, in which checkpoints from a fine-tuning version stream are retained as candidate domain specialists; 2. a lightweight image-level domain router that dispatches pages to experts, with an ex- plicit leakage-safe training protocol, a source- balancing fix for visually adjacent domains, and a domain-label oracle; 3. empirical evidence on historical Manchu that routing matches the best evaluated do- main scores at two-decimal precision (regu- lar 0.30%, memorial 1.57%, running 4.83% CER), which no single checkpoint attains, us- ing two checkpoints not trained specifically for their selected domains and one purpose- trained running-script checkpoint; and 4. a reproducible evaluation package consisting of the model assignments, router protocol, source-separated test sets, and per-page pre- dictions. 2 Background The Manchu script and its romanization. Manchu is written in an alphabetic script adapted, by way of Mongolian, from Old Uyghur; it runs vertically from top to bottom, with columns or- dered left to right. It is a connected script: let- ters join into a continuous word-stem, and most letters take distinct initial, medial, and final forms according to position, with further contextual vari- ants. The standard scholarly transcription is the M ̈ ollendorff romanization, which maps the script to a Latin alphabet with a few diacritics (e.g., ˇ s, ̄ u, ˇ z). Following common practice and the prior benchmark [5], our system outputs M ̈ ollendorff ro- manization rather than Manchu Unicode: roman- ization is what most Manchu scholarship reads and searches, and it sidesteps the still-unsettled digital rendering of the script. The recognition target is therefore a Latin character string, and our metric is a character error rate over it. The written record and its visual styles. Sur- viving Manchu materials were produced using different technologies and for different purposes. Manchu palaeography therefore distinguishes sev- eral writing styles rather than one uniform hand; the First Historical Archives of China groups much of its material into categories such as memorial- petition, book, proclamation, running-hand, and woodblock styles. For recognition we group our corpus into three visual styles: 2 • regular script—the upright, well-separated letterforms of woodblock and movable-type imprints (the archives’ woodblock style) and of formal book hands (its book style): reign chronicles, primers, printed collections, and neatly written archival pages; • memorial script—the controlled semi-cursive chancery hand of palace memorials submitted to the throne (the archives’ memorial-petition style), frequently bearing vermilion imperial rescripts; • running script—the fluent, ligature-heavy semi-cursive of manuscripts and correspon- dence (the archives’ running-hand style). Although these materials use the same language and script, their letterforms and page layouts dif- fer substantially. Our corpus—the Old Manchu Archives, the manc.hu manuscript database, printed collections,and palace memorials— therefore spans all three styles rather than a single clean source. 3 Related Work OCR for ancient and historical scripts. Recent work on OCR and HTR for historical and low- resource scripts addresses similar data constraints. Examples include cuneiform sign detection with annotated corpora and processing pipelines [4]; new datasets for East Syriac [26] and Sahidic Cop- tic [31]; recognition of Ge’ez/Ethiopic [2]; Dead Sea Scroll palaeography [29]; and Arabic-script re- sources using open corpora [28] and multimodal models [39]. OCR for Asian and minority scripts. Re- lated work covers palm-leaf datasets [18], In- dic layout parsing [30], Old Uyghur transcription with vision–language fine-tuning [8], and routed adapters for Tibetan, Yi, Shui, and Dongba [23]. Broader benchmarks continue to show substantial variation in performance across scripts [17]. Methods for the low-resource regime. Com- mon strategies include transfer learning and parameter-efficient fine-tuning of pretrained back- bones, such as TrOCR [22]; synthetic data gen- eration; few-shot learning from glyph exemplars [37]; and data augmentation [32]. Most systems are organized around one model per language, col- lection, or hand [27]. Our setting differs in that one language contains several visual domains, and the training history itself supplies a pool of check- points with different domain strengths. ManchuOCR. Machine-learningworkon Manchu OCR is recent, and existing systems are mostly word-level. One line of work segments pages into word images—using projection profiles or stroke growth—and recognizes each word with a CRNN [35], a sliding-window CNN [42], or related models; ManchuOCR is an open-source example [38].More recent work fine-tunes vision–language models on synthetic word images and evaluates them at the word level [5]. Institu- tional recognizers also exist but are not publicly available; as far as we understand, they likewise use a segment-then-recognize design. Document image-to-sequence models. Page- level recognizers avoid an explicit segmenta- tion stage by encoding a page image and emit- ting the transcription directly.Donut [20] and Nougat [3] are examples of document encoder– decoder models; Pix2Struct [21], Kosmos-2.5 [25], and GOT-OCR2.0 [40] use related document or screenshot-reading objectives.General-purpose vision–language models [1, 24] can also transcribe pages, although specialized document models re- main useful for high-precision settings.Line- and word-level recognizers, including CRNN/CTC [35], TrOCR [22], and kraken [19], require an up- stream segmentation stage that page-level models avoid. Mixtures of experts, ensembles, and reusing checkpoints. Routing inputs to specialized mod- els is a classical idea [14], and modern mixture- of-experts systems learn routers and experts jointly [34, 10, 33, 15, 7]. Our system operates at the model level: the experts are independently trained checkpoints and the router is an external classifier. Snapshot ensembles retain several checkpoints from one training run [12], whereas model soups and stochastic weight averaging merge model pa- rameters [41, 13]. These methods generally target one data distribution. We instead use checkpoints with different domain scores and select one model per page, so the comparison concerns domain spe- cialization rather than output averaging or weight 3 merging. 4 Task and Evaluation Protocol The task is image-to-text transcription of a com- plete Manchu page or folio into M ̈ ollendorff-style Romanization. Input images follow a fixed con- vention recorded with the data, including image enhancement and optional cropping or denoising. The original orientation and reading order are pre- served for evaluation and human review. For a page with reference string g and prediction p, character error rate is CER(p, g) = Lev(p, g) |g| ,(1) with whitespace normalized consistently and the page evaluated in its original reading order. We report the mean over pages. This is a page-level CER, not a word accuracy, and must not be com- pared directly with word-image benchmarks (Sec- tion 9). Frozen, source-separated test sets. Test report- ing uses three domain-specific sets held fixed across the project so that successive versions re- main comparable: • regular script: 125 pages of regular, well- separated letterforms from several sources (woodblock and movable-type collections and the regular hand of the Old Manchu Archives), the principal regular-script test set; • memorial: 10 handwritten palace-memorial pages (a memorial-domain probe); • running script: 13 semi-cursive manuscript pages (a running-script probe). Each set is derived directly from the test split of the master data registry by a single construc- tion script, so it remains synchronized with the frozen split and is never drawn from a stale snap- shot.The handwriting sets are small and are treated as domain-shift probes rather than large- sample estimates; enlarging them with additional expert-transcribed held-out pages is ongoing par- allel work, kept separate from these frozen sets to preserve comparability. Table 1: Page-level CER (%) of representative Nougat checkpoints on the three frozen test sets (regular P125, memorial Z10, running C13). Bold marks each test set’s best result (column mini- mum). No single version wins every column. The upper group contains checkpoints selected after training; the lower group contains purpose-trained v5.0 variants. VersionRegularMemorialRunning checkpoints selected after training v3.00.303.4047.09 v3.0 (seed 2)0.301.5751.59 v4.1 (seed 2)0.453.6526.68 deliberate (purpose-trained v5.0) v5.0 (mem.-tuned)0.411.606.86 v5.0 (curs.-tuned)0.793.294.83 5 Domains and the Version Stream The project maintains a source-aware data registry (about 9,054 page/folio records across nine sources as of version v4.1, and still growing) with a per- manent train/validation/test split; the three test sets above are frozen subsets and are not used for model training or router checkpoint selection. They are, however, used retrospectively to designate the ex- pert pool after the training runs are complete, as disclosed in Section 6. Successive training ver- sions differ in their data snapshot and augmenta- tion recipe. In particular, regular-script sources use mild augmentation, whereas the memorial and running-script sources receive stronger, style- specific augmentation and source-aware oversam- pling, so that later versions increasingly favor the harder handwriting styles. The key empirical observation is that no single version is best on all three test sets. Table 1 reports page-level CER for representative Nougat check- points. The strongest regular-script and memo- rial checkpoints are v3.0 and v3.0 s2, respectively; neither was trained specifically for the final as- signment. The best earlier checkpoint on running script, v4.1 s2, reaches only 26.68%. A later v5.0 run trained with running-script data reduces this error to 4.83%, but its regular-script and memo- rial scores remain above those of the earlier check- points. These results provide the basis for selecting different checkpoints for different domains. 4 0.30.40.50.60.70.80.9 Regular-script CER (%) → worse 0 10 20 30 40 50 Cursive CER (%) → worse v3.0 v3.0-s2 v4.1-s2 v5.0-qam v5.0-curs coordinate-wise minimum Checkpoint trade-off projection 2D trade-off envelope selected after training purpose-trained (v5.0) coordinate-wise minimum Figure 1: A two-dimensional projection of the checkpoint trade-off: regular-script CER versus running-script CER (from Table 1). Grey points are checkpoints selected after training; blue points are purpose-trained v5.0 variants. The red star marks the coordinate-wise minimum on the two displayed test sets. The plot does not characterize the full three-set error space. A benchmark-scoped trade-off view. Figure 1 projects errors on two of the three test sets, with lower CER indicating better performance. It vi- sualizes the observed checkpoint trade-off rather than establishing non-dominance in the full three- dimensional space. The coordinate-wise minima in this pool are (0.30, 1.57, 4.83) for regular, memo- rial, and running script, respectively. No single checkpoint attains all three values. The domain- label oracle combines the corresponding special- ists, and the learned router matches it at two- decimal precision. The result is specific to the checkpoints and test sets used here and should be reassessed as the data and model pool expand. 6 Method: Version Experts and a Domain Router Problem statement. Figure 2 sketches the routed system. Let a page image x belong to one of K visual domains, and lete 1 , . . . , e M be a pool of transcription experts, each a complete page-to- text model. A router r maps x to an expert index, and the system transcribes x as e r(x) (x). Writing c(e, x) for the CER of expert e on page x, we dis- tinguish two reference policies. The domain-label oracle uses the true domain label y(x) and dis- patches to its designated expert e y(x) ; this is the oracle reported in our experiments. A stronger per- page oracle would choose arg min m c(e m , x) and is only a theoretical upper bound, not an experi- mental baseline here. With one designated spe- cialist per domain, the learned router equals the domain-label oracle when its domain prediction is correct, but it need not equal the per-page oracle. The system therefore reduces to domain prediction followed by expert selection. Since the experts are existing checkpoints, the added model is only the page-level classifier. Expertpool. Allexpertsshareone architecture—aNougatdocumentencoder– decoder [3] adapted to Manchu Romanization— and differ in their training version. The appendix reports a common-recipe comparison of three candidate backbones; Nougat performed most consistently across the three styles. From this pool we designate three domain experts, each the best available checkpoint on its domain in Table 1: v3.0 for regular script, v3.0 s2 for memorials, and v5.0 for running script. The first two checkpoints were selected after training rather than trained for those final assignments; the v5.0 running-script checkpoint was trained specifically for that do- main. Because the expert pool was designated retrospectively from the completed test-set results, the routed number is a frozen-benchmark compari- son rather than an independently held-out estimate of expert-pool selection. Domain router. The router is a small resid- ual image classifier [11] (ResNet-18, ImageNet- initialized, 224×224 input) that maps a page image to a domain label (regular / running / memorial). At inference the page is transcribed by the expert cor- responding to the predicted label. Because read- ing a low-resolution page is enough to judge visual style, the router is small and fast; it adds a single classifier forward pass per page. Source-balanced training. The router’s super- vision is imbalanced: the training split contains roughly 8000 regular-script, 153 running-script, and 50 memorial pages. Inverse-frequency class weighting and weighted sampling address the class imbalance. A second imbalance occurs within the regular-script class, which is dominated by one archival source. A naive router underfits the vi- sually distinct manc.hu woodblock print and con- 5 PAGE-LEVEL INFERENCE page image DOMAIN ROUTER ResNet-18 style label regular · memorial · running CHECKPOINT-DERIVED EXPERTS the router dispatches one page to one expert REGULAR expert checkpoint · v3.0 MEMORIAL expert checkpoint · v3.0 s2 RUNNING expert checkpoint · v5.0 MÖLLENDORFF transcription selected pathidle alternatives Figure 2: The routed system. A lightweight domain router (a small image classifier) reads each page and predicts its visual style; the page is then transcribed by the corresponding domain expert—a Nougat checkpoint drawn from the project’s own version stream. The solid orange path shows a memorial page routed to the memorial expert (v3.0 s2); the dashed gray branches are alternative experts left idle for that page. fuses it with running-hand pages from the same collection. We therefore assign equal sampling mass to regular-script sources. In the ablation, this change reduces the resulting confusion between vi- sually adjacent sources. Leakage-safe training. The router is trained and selected only on training-split pages; the frozen test sets (regular, memorial, running) are not used for router fitting or router checkpoint selection. Domain labels are derived from source meta- data (running-script source → running; memorial source → memorial; all others → regular). Be- cause project policy removes the memorial source from the global validation pool, we reserve a small held-out slice of training memorial pages purely to monitor memorial recall during router check- point selection; these pages are never part of the memorial test set. Separately, the expert pool itself was designated retrospectively from the completed frozen-test-set results, so this experiment does not claim a fully blind estimate of model-selection per- formance. Domain-label oracle. To separate router error from expert quality, we report a domain-label ora- cle that uses each page’s true domain label and the fixed domain-to-expert mapping. The gap between the learned router and this oracle measures the cost of domain-routing mistakes; it is distinct from the stronger per-page oracle defined above. 7 Experiments We evaluate on the three frozen test sets. For ev- ery page we run each expert, record its CER, and then report, for each set, every single expert ap- plied to the full set, the domain-label oracle, and the learned router. Routing matches the domain-label oracle at re- ported precision. With three experts—regular script (v3.0), running script (v5.0), and memo- rial (v3.0 s2)—the router dispatches pages to the corresponding specialist. It matches the domain- label oracle to two decimal places on all three test sets: 0.30% on regular script, 1.57% on memori- als, and 4.83% on running script (Table 2). Its do- main accuracy is 99.3% (147/148; Figure 3). The single-expert results show why selection matters: the regular-script expert reaches 47.1% CER on running script, whereas the running-script expert reaches 0.79% on regular script. Effect of domain-specific training. The com- parison separates the contribution of the check- point history from that of the purpose-trained ex- pert. The best earlier checkpoint scores 26.68% on running script; the running-script v5.0 checkpoint reduces this to 4.83%. The same v5.0 campaign 6 Table 2: Three-expert routing (regular v3.0 + run- ning v5.0 + memorial v3.0 s2), page-level CER (%) on the frozen test sets (regular P125, memo- rial Z10, running C13). The router matches the domain-label oracle to two decimal places on all three sets. Domain accuracy is 99.3% (147/148). Off-diagonal columns show the cost of sending a page to the wrong expert. Test set reg only run only mem onlyoracleroutermis regular0.300.790.300.300.301/125 memorial3.403.291.571.571.570/10 running47.094.8351.594.834.830/13 regularrunningmemorial routed to (predicted domain) regular running memorial true domain 124 1· · 13 · · 10 router accuracy 147/148 = 99.3% Figure 3: Domain-router confusion matrix on the frozen test sets (rows: true domain; columns: ex- pert routed to). The learned router places 147 of 148 pages correctly (99.3%); the single error is one atypical manc.hu regular-script page sent to the running-script expert. Its regular-script CER is 0.299%, versus 0.296% for the domain-label or- acle, which rounds to 0.30% for both. does not replace the selected checkpoints on regu- lar script or memorials: its best scores on those test sets are 0.41% and 1.60%, compared with 0.30% and 1.57%. The selected pool therefore combines checkpoints from different stages of training rather than using a single final version. Router errors and source balancing. One page is misrouted: an atypical manc.hu regular-script page is sent to the running-script expert. The ef- fect on the regular-script test set is small (0.299% routed CER versus 0.296% for the domain-label oracle; both round to 0.30%). The ablation shows a larger failure mode when regular-script sources other regular manc.hu regular memorialrunning 0 2 4 6 routed page CER (%) 1/113 0/113 7/12 1/12 0/100/10 0/130/13 annotations = misrouted pages / total; orange meeting the dashed line = router matches oracle oracle (best expert)naive samplingsource-balanced Figure 4: Source-balanced router training (abla- tion) on the frozen test sets. Within the regular- script class, a naive router underfits the minority manc.hu woodblock print and misroutes most of it to the running-script expert (blue, 7/12), inflating its routed CER to 4.1% against a domain-label ora- cle of 0.5%; balancing the regular-script sources (orange) recovers domain-label-oracle-level rout- ing (1/12, 0.6%). The dominant archival regular hand and the other styles are unaffected. are not balanced: the router confuses manc.hu woodblock pages with running-hand pages from the same collection (Figure 4). Equalizing the sam- pling mass of regular-script sources removes most of this confusion. Why a shared Nougat backbone. All experts share the Nougat encoder–decoder. Under a com- mon recipe, a from-scratch ResNet-50 recognizer is comparable to Nougat on regular script (0.48% versus 0.56%) but performs poorly on handwrit- ing (5.04% on memorials and 18.67% on running script). Donut (Swin + mBART) is weaker on reg- ular script (1.23%) but reaches 5.89% on running script (Appendix A). These results support using Nougat for the pooled experts. 8 Discussion The benefit of routing is largest when domain- specific experts differ substantially. On the regular- script test set, routing to v3.0 gives the same rounded CER as applying that expert to every page. On running script, the purpose-trained expert re- duces CER without changing the selected regular- script and memorial checkpoints. This separation allows a new domain to be added as an expert and a router class without replacing the existing recog- nizers. Within this benchmark, a new specialist can lower the routed error when it improves a domain 7 not already covered by the selected pool. Future work includes stronger running-script models, spe- cialists for seals or marginalia, and larger handwrit- ing test sets so that these comparisons can be esti- mated with greater precision. 9 Relation to Prior Manchu OCR Results The most relevant prior system fine-tunes a vision– language model on synthetic word images and evaluates at the word level on regular script [5]. An open-source predecessor, ManchuOCR [38]— a community code release without an accompa- nying paper—likewise operates at the word level, segmenting a page into word images and recogniz- ing each. Because our system instead transcribes complete pages across three visual styles, neither is directly comparable to it. (At the word level on regular script, our page-to-page system reaches 98.86% word accuracy, for reference.) These lines of work are complementary; a shared public word- and-page benchmark would allow comparison on common data. 10 Limitations and Release Status The handwriting test sets are small (the memo- rial and running-script sets have 10 and 13 pages), so they are domain-shift probes rather than large- sample estimates; enlarging them with additional expert-transcribed held-out pages is ongoing. The experts come from successive data snapshots rather than a single controlled recipe sweep, so version differences confound data and augmentation. The router uses a coarse three-way domain label; finer style distinctions may require more classes and more monitoring data, especially for domains re- moved from the validation pool. Finally, this is a proof of concept with a small number of experts; scaling the pool and the router to more domains is future work. Further progress depends on collaboration with fluent Manchu readers. In particular, expert review of difficult pages, correction of transcription errors, and expansion of running-script annotations would help improve recognition accuracy where the cur- rent system remains weakest. We also welcome collaborators who would like to join our team. Our broader work focuses on natural-language processing for low-resource lan- guages, not only Manchu, with the aim of help- ing preserve cultural heritage. Previous outputs have addressed Syriac, Latin, Greek, and Manchu. We also welcome contact from researchers and cultural-heritage projects working with other low- resource languages and related NLP needs. The release plan has four stages: (1) freeze the split and preprocessing manifest; (2) publish the router, evaluation scripts, and per-page predic- tion files; (3) publish expert weights subject to the underlying data and model licenses; and (4) add a public, independently adjudicated handwriting benchmark. This arXiv version reports the current state and does not imply that the later stages are complete. A public trial interface for the current page-level OCR is available at http://124. 223.33.6/manchu-ocr/. Translation is be- ing developed as a separate extension and is not evaluated in this paper. As a next-version extension, we are exploring a three-checkpoint ensemble within each visual do- main: v3.0, res50-print, and Donut for regular script; qam-s42, res50-bal, and Donut for memo- rials; and curs-s43, res50-print, and Donut for run- ning script. The current design uses anchored vot- ing for regular script and memorials and ROVER decoding for running script, with provisional target CERs of 0.28%, 1.39%, and 3.81%, respectively. These targets are not part of the present benchmark results; they will be formally evaluated under the frozen P125/Z10/C13 protocol in a future version. A companion Manchu-language parser is availableat http://124.223.33.6/ manchu-parser/.It currently supports parsing,translation,andword-formanal- ysis.The first four integrated dictionaries come from the Tohoku University collection at http://hkuri.cneas.tohoku.ac.jp/ project1/manchu/list?groupId=11, with additional data from https://manc.hu/. The parser is a separate language resource and is not part of the OCR benchmark reported here. 11 Conclusion We presented a multi-expert routing approach for multi-domain low-resource OCR. The system reuses checkpoints from an iterative fine-tuning process as domain experts and selects among them with a page-level visual classifier. On the frozen 8 Manchu test sets, it matches the best evaluated do- main scores at two-decimal precision: 0.30% on regular script, 1.57% on memorials, and 4.83% on running script. The result supports retaining complementary checkpoints and adding a special- ist when a domain remains underrepresented. The paper specifies the evaluation protocol and routing analysis; the release of models and data remains subject to the applicable licenses. References [1] Jinze Bai, Shuai Bai, Shusheng Yang, Shi- jie Wang, Sinan Tan, Peng Wang, Jun- yang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. [2] Birhanu Hailu Belay, Isabelle Guyon, Tadele Mengiste, Bezawork Tilahun, Marcus Li- wicki, Tesfa Tegegne, Romain Egele, and Tsiyon Worku.HHD-Ethiopic: A histor- ical handwritten dataset for Ethiopic OCR with baseline models and human-level perfor- mance. In Document Analysis and Recogni- tion – ICDAR 2024, pages 23–38. Springer, 2024. [3] Lukas Blecher, Guillem Cucurull, Thomas Scialom, and Robert Stojnic. Nougat: Neu- ral optical understanding for academic doc- uments.arXiv preprint arXiv:2308.13418, 2023. [4] Wentao Che, Esteban Garc ́ es Arias, Asim Niaz, Andreas Bender, and Enrique Jim ́ enez. Automated sign detection across the Elec- tronic Babylonian Library:A large-scale dataset and end-to-end cuneiform OCR pipeline. arXiv preprint arXiv:2606.22608, 2026. [5] Yan Hon Michael Chung and Donghyeok Choi.Finetuning vision-language models as ocr systems for low-resource languages: A case study of manchu.arXiv preprint arXiv:2507.06761, 2025. [6] Pamela Kyle Crossley and Evelyn S. Rawski. A profile of the Manchu language in ch’ing history. Harvard Journal of Asiatic Studies, 53(1):63–102, 1993. [7] Damai Dai, Chengqi Deng, Chenggang Zhao, R. X. Xu, Huazuo Gao, Deli Chen, Jiashi Li, Wangding Zeng, Xingkai Yu, Y. Wu, et al. DeepSeekMoE: Towards ultimate expert spe- cialization in mixture-of-experts language models.arXiv preprint arXiv:2401.06066, 2024. [8] Mehmet Oguz Derin and Erdem Uc ̧ar. Old Uyghur OCR: The first work-in-progress via reproducing fine-tuning of VLMs. Journal of Old Turkic Studies, 9(2):313–318, 2025. [9] Mark C. Elliott.The Manchu-language archives of the Qing dynasty and the origins of the palace memorial system. Late Imperial China, 22(1):1–70, 2001. [10] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to tril- lion parameter models with simple and effi- cient sparsity. Journal of Machine Learning Research, 23(120):1–39, 2022. [11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for im- age recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–778, 2016. [12] Gao Huang, Yixuan Li, Geoff Pleiss, Zhuang Liu, John E. Hopcroft, and Kilian Q. Wein- berger.Snapshot ensembles: Train 1, get M for free. In International Conference on Learning Representations (ICLR), 2017. [13] Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization. In Confer- ence on Uncertainty in Artificial Intelligence (UAI), 2018. [14] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural Computa- tion, 3(1):79–87, 1991. [15] Albert Q. Jiang, Alexandre Sablayrolles, An- toine Roux, et al. Mixtral of experts. arXiv preprint arXiv:2401.04088, 2024. [16] PhilipKahle,SebastianColutto, G ̈ unter Hackl, and G ̈ unter M ̈ uhlberger. 9 Transkribus—a service platform for tran- scription,recognitionandretrievalof historical documents. In 14th IAPR Inter- national Conference on Document Analysis and Recognition (ICDAR), volume 4, pages 19–24. IEEE, 2017. [17] Amir Hossein Kargaran, Nafiseh Nikeghbal, Jana Diesner, Franc ̧ois Yvon, and Hinrich Sch ̈ utze. GlotOCR Bench: OCR models still struggle beyond a handful of Unicode scripts. arXiv preprint arXiv:2604.12978, 2026. [18] Made Windu Antara Kesiman, Sophea Prum, Jean-Christophe Burie, and Jean-Marc Ogier. AMADI LontarSet:The first handwritten Balinese palm leaf manuscripts dataset. In 15th International Conference on Frontiers in Handwriting Recognition (ICFHR), pages 168–173. IEEE, 2016. [19] Benjamin Kiessling. Kraken—an universal text recognizer for the humanities. In Digital Humanities Conference (DH2019), 2019. [20] Geewook Kim, Teakgyu Hong, Moon- bin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sang- doo Yun, Dongyoon Han, and Seunghyun Park.Ocr-free document understanding transformer.In European Conference on Computer Vision, pages 498–517. Springer, 2022. [21] Kenton Lee, Mandar Joshi, Iulia Raluca Turc, Hexiang Hu, Fangyu Liu, Julian Mar- tin Eisenschlos, Urvashi Khandelwal, Pe- ter Shaw, Ming-Wei Chang, and Kristina Toutanova. Pix2Struct: Screenshot parsing as pretraining for visual language understand- ing. In International Conference on Machine Learning (ICML), pages 18893–18912, 2023. [22] Minghao Li, Tengchao Lv, Jingye Chen, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, and Furu Wei. TrOCR: Transformer-based optical character recogni- tion with pre-trained models. In Proceed- ings of the AAAI Conference on Artificial In- telligence, volume 37, pages 13094–13102, 2023. [23] Bonan Liu, Zeyu Zhang, Bingbing Meng, Han Wang, Hanshuo Zhang, Chengping Wang, Daji Ergu, and Ying Cai.Om- niOCR: Generalist OCR for ethnic minority languages. arXiv preprint arXiv:2602.21042, 2026. [24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tun- ing. Advances in neural information process- ing systems, 36, 2024. [25] Tengchao Lv, Yupan Huang, Jingye Chen, Lei Cui, et al.Kosmos-2.5:A mul- timodal literate model.arXiv preprint arXiv:2309.11419, 2023. [26] Ameer Majeed and Hossein Hassani. Ancient but digitized: Developing handwritten optical character recognition for East Syriac script through creating KHAMIS dataset.arXiv preprint arXiv:2408.13631, 2024. [27] Konstantina Nikolaidou, Mathias Seuret, Hamam Mokayed, and Marcus Liwicki. A survey of historical document image datasets. International Journal on Document Analy- sis and Recognition (IJDAR), 25(4):305–338, 2022. [28] Open Islamicate Texts Initiative.OpenITI MAKHZAN: An open annotated dataset of Arabic, Persian, Ottoman Turkish, and Urdu print and manuscript data.Zenodo, 2025. dataset. [29] Mladen Popovi ́ c, Maruf A. Dhali, and Lam- bert Schomaker. Artificial intelligence based writer identification generates new evidence for the unknown scribes of the Dead Sea Scrolls exemplified by the great Isaiah scroll (1qisaa). PLOS ONE, 16(4):e0249769, 2021. [30] Abhishek Prusty, Sowmya Aitha, Abhishek Trivedi, and Ravi Kiran Sarvadevabhatla. Indiscapes:Instance segmentation net- works for layout parsing of historical Indic manuscripts.In International Conference on Document Analysis and Recognition (IC- DAR), pages 999–1006. IEEE, 2019. [31] FabioQuattrini,CarmineZaccagnino, Costanza Bianchi, Silvia Cascianelli, and Rita Cucchiara. A text recognition dataset from Sahidic Coptic ancient manuscripts. arXiv preprint arXiv:2606.15987, 2026. 10 [32] Yassin Hussein Rassul, Aram M. Ahmed, Polla Fattah, Bryar A. Hassan, Arwaa W. Ab- dulkareem, Tarik A. Rashid, and Joan Lu. Advancing offline handwritten text recog- nition: A systematic review of data aug- mentation and generation techniques. arXiv preprint arXiv:2507.06275, 2025. [33] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenat- ton, Andr ́ e Susano Pinto, Daniel Keysers, and Neil Houlsby.Scaling vision with sparse mixture of experts. In Advances in Neural Information Processing Systems (NeurIPS), 2021. [34] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture- of-experts layer. In International Conference on Learning Representations (ICLR), 2017. [35] Baoguang Shi, Xiang Bai, and Cong Yao. An end-to-end trainable neural network for image-based sequence recognition and its ap- plication to scene text recognition.IEEE transactions on pattern analysis and machine intelligence, 39(11):2298–2304, 2016. [36] Ray Smith. An overview of the Tesseract OCR engine. In Ninth International Confer- ence on Document Analysis and Recognition (ICDAR), volume 2, pages 629–633. IEEE, 2007. [37] Mohamed Ali Souibgui, Alicia Forn ́ es, Yousri Kessentini, and Be ́ ata Megyesi. Few shots are all you need: A progressive learn- ing approach for low resource handwritten text recognition. Pattern Recognition Letters, 160:43–49, 2022. [38] tyotakuki.ManchuOCR: An OCR system for the Manchu script. https://github. com/tyotakuki/ManchuOCR.GitHub repository, accessed 2026-07-14. [39] Ahmed Wasfy, Omer Nacar, Abdelakreem Elkhateb, Mahmoud Reda, Omar Elshehy, Adel Ammar, and Wadii Boulila.QARI- OCR: High-fidelity Arabic text recognition through multimodal large language model adaptation. arXiv preprint arXiv:2506.02295, 2025. [40] Haoran Wei, Chenglong Liu, Jinyue Chen, et al. General OCR theory: Towards OCR- 2.0 via a unified end-to-end model. arXiv preprint arXiv:2409.01704, 2024. [41] MitchellWortsman,GabrielIlharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S. Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, and Ludwig Schmidt. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International Conference on Machine Learning (ICML), pages 23965–23998, 2022. [42] Diandian Zhang, Yan Liu, Zhuowei Wang, and Depei Wang.OCR with the deep CNN model for ligature script-based lan- guages like Manchu. Scientific Programming, 2021:5520338, 2021. 11 A Architecture comparison Table 3 compares three page-level backbones trained with the same source-balanced (v5.0) recipe on the frozen test sets. Nougat is the most balanced of the three architectures: the from-scratch ResNet-50 recognizer is strong on regular script but performs poorly on handwriting, while Donut is weaker on regular script. This comparison motivates the use of Nougat for all experts in the routed pool. Table 3: Page-level CER (%) of three backbones under a common source-balanced recipe (regular P125, memorial Z10, running C13). Nougat balances all three styles; ResNet-50 is a regular-script specialist; Donut trails on regular script. BackboneRegularMemorialRunning Nougat (enc–dec)0.562.335.92 ResNet-50 (from scratch)0.485.0418.67 Donut (Swin + mBART)1.233.755.89 12 B Qualitative examples Figure 5 shows the routed system’s output on one held-out page from each of the three visual styles, in M ̈ ollendorff romanization. The regular-script and running-script pages are transcribed exactly; the memorial page carries a small residual error rate. These pages are drawn from the frozen test sets and were never seen in training. Regular script manc.hu regular hand page CER exact OBURE BE SARA TEILE AKŪ TONDO OJORONGGE GELI HIYOOŠUN CI BANJINARA BE SACI OMBIKAI UBABE GUCUSE ULHICI ... Memorial script NPM palace memorial no. 156603 page CER 0.61% JERGI AHASI CI ANAME GELI AKŪNGGE URSE OCI EICI INI AHŪN DEO NIYAMAN HŪNCIHIN DOLO CIHANGGAI AISILAME ... Running script manc.hu manuscript mh5110 page CER exact JUWAN CUWAN I BELE UNGGIHE MIMBE ALIYA SEHEBI JE MA TANG NI BADE JUWAN EMU INENGGI INDEHE JUWAN JUWEDE ... Model output on one held-out page per visual style (Möllendorff romanization) Figure 5: Model output on one held-out page per visual style (regular, memorial, running), with the page-level CER and the opening of the predicted transcription. Live showcase: http://124.223. 33.6/manchu-ocr/. 13 C Parser showcase In addition to the OCR interface, the project provides a public Manchu-language assistance parser at http://124.223.33.6/manchu-parser/.The tool supports parsing, translation, and full word-form analysis.Its first four integrated Manchu dictionaries come from the Tohoku University collection at http://hkuri.cneas.tohoku.ac.jp/project1/manchu/list? groupId=11, with additional data from https://manc.hu/. The example below illustrates the current interface: dictionary glosses are retained in their original Chinese/Japanese form, while the En- glish translation and retrieval-augmented morphological analysis remain experimental. Figure 6: Example output from the public Manchu-language assistance parser. The interface shows parsing, Chinese translation, an experimental English translation, dictionary hits, and inline morpho- logical segmentation for one sentence. The displayed “10/10” match is an interface example, not an independent accuracy evaluation. 14

Instant research discovery