Beyond Recognition: Compact Multi-Domain Arabic Manuscript HTR with Candidate-Selection Analysis and Evidence-Preserving Review

Abdullah Ahmed Ali; Mohammed Thamer Abdulhadi; Ali Haider Safaa; Dhulfiqar Mahdi Wadi

cs.CVAug 19, 2026

Beyond Recognition: Compact Multi-Domain Arabic Manuscript HTR with Candidate-Selection Analysis and Evidence-Preserving Review

Authors: Abdullah Ahmed Ali, Mohammed Thamer Abdulhadi, Ali Haider Safaa, Dhulfiqar Mahdi Wadi

Organizations: University of Technology, Baghdad, Iraq

Abstract

Historical Arabic manuscript transcription is not only a recognition problem. A usable scholarly system must cope with shifting hands and layouts, preserve uncertain readings, distinguish visual evidence from linguistic plausibility, and record the researcher's final decision. We present Phoenix, a 4.99-million-parameter CNN-BiLSTM-CTC recognizer, and Athar, an evidence-aware review workflow built around it. Phoenix is adapted across archival, Maghrebi, and historical manuscript domains using document-aware replay, an expanded 81-symbol codec, and forgetting guards that reject checkpoints that improve a new domain at unacceptable cost to previous domains. In a pre-specified held-out comparison against the preceding checkpoint, frozen before evaluation and scored with greedy decoding and raw references, Phoenix reduced CER from 22.12% to 17.86% on 10,594 Agapet lines and from 17.72% to 11.84% on 11,684 Omar lines, while regressing from 10.39% to 10.72% on 164 TariMa lines. Across the two large held-out sets, character-weighted CER fell from 19.98% to 14.93%, a 25.3% relative error reduction. A separate same-protocol development diagnostic found the lowest CER for Phoenix on four of four comparable domains (9.59% unweighted macro CER). An N-best diagnostic revealed a 2.15-point oracle gap between beam decoding and Oracle@25, while neural text rerankers, consensus MBR, CTC-posterior quality estimation, and local pre-CTC hidden-state quality estimation recovered less than 4% of this gap. Athar therefore preserves the visual reading, exposes bounded alternatives, uses local language models conservatively, retrieves source parallels with unique, ambiguous, or abstain states, and exports auditable TEI and PAGE-XML records. The results support evaluating manuscript HTR as auditable evidence management rather than silent text replacement.

Figures & tables

Figure 1: The Athar architecture. Evidence channels are additive and labelled; they do not overwrite the visual reading or each other.

Tier	Permitted interpretation
Sealed evaluation	Frozen exp8 vs. exp9; raw references; greedy decoding; no tuning after opening. Primary release evidence.
Development diagnostic	Identical line inputs and protocol across models; some guards influenced selection. Domain diagnosis, not final generalization.
Historical / server diagnostic	Aggregated diagnostic result preserved from historical development records; informative for decoding analysis, not primary held-out generalization.
Additional server benchmark	Owner-run Phoenix–Baseer aggregates; incomplete reproducibility bundle. Preliminary comparison only.

Table 1: Evidence tiers used in this paper.

Dataset	Domain	Training/adaptation role	Separation unit	Evaluation role
Muharaf	19th–20th c. archival handwriting	Replay and archival adaptation	Document	Development guard; not a new sealed result
RASAM	Maghrebi manuscripts	Core training and replay	Manuscript/page manifests	Development guard and historical results
TariMa	Diverse Maghrebi/historical hands	Replay	Manuscript where available	164-line sealed release set
Agapet	13th–17th c. Christian Arabic	SA-418 adaptation	Manuscript	Sin423 + BnF Arabe 76 sealed set
Omar	1951–1965 archival memoirs	Six-document adaptation side	Document	Eleven-document sealed set

Table 2: Dataset roles in the current study. Exact training counts depend on replay sampling; the table records the unit of separation and experimental role.

Set	Lines	exp8 CER	exp9 CER	$\Delta$ CER	exp9 WER	Word acc. exp8 $\rightarrow$ exp9
Agapet	10,594	22.124	17.861	$-4.263$	58.798	23.91 $\rightarrow$ 33.23
Omar	11,684	17.718	11.835	$-5.883$	42.917	33.40 $\rightarrow$ 46.83
TariMa	164	10.386	10.718	$+0.332$	38.886	51.32 $\rightarrow$ 49.23

Table 3: Sealed release comparison. Raw references, greedy decoding, no language model. Lower CER/WER is better; higher word accuracy is better.

Figure 2: CER on the three sealed release sets. The TariMa regression is included in the release decision.

Guard	Lines	Original Muharaf	exp6	exp8	Phoenix / exp9
Agapet	991	33.15 / 16.62	23.82 / 24.71	20.08 / 33.58	9.47 / 67.40
Omar	1,143	9.26 / 60.08	26.88 / 21.25	12.06 / 53.90	8.91 / 63.95
RASAM	1,789	39.01 / 12.06	9.04 / 66.78	7.95 / 69.90	7.67 / 70.46
Muharaf	920	13.28 / 64.15	36.47 / 15.74	12.72 / 58.30	12.31 / 59.32
Unweighted macro	4,843	23.68 / 38.23	24.05 / 32.12	13.20 / 53.92	9.59 / 65.28

Table 4: Same-protocol development diagnostic. Each cell is CER / word accuracy (%). Lowest CER is bold.

Figure 3: Controlled development CER. The figure is diagnostic, not a sealed leaderboard.

Decoder	Norm. CER	Norm. WER	Exact lines
Greedy	11.52	38.20	10.5
Archive LM	11.04	32.09	14.6
Mixed LM	10.89	31.88	14.8
Mismatched general LM	14.78	44.98	5.1

Table 5: Domain-conditioned LM diagnostic on 920 archival development lines.

Metric	Phoenix	Baseer__Nakba	Observed difference
Macro CER, 4 domains (incl. task-matched Omar)	9.71%	28.67%	$-18.96$ points
Macro CER, 3 domains (non-Omar, descriptive)	10.65%	38.07%	$-27.42$ points
WER (aggregation denominator not supplied)	34.37%	59.52%	$-25.15$ points
Exact-line accuracy	13.38%	26.38%	Baseer $+13.00$ points
Parameters	4.99M	$\approx$ 3.75B	Phoenix $\approx 752\times$ smaller
Model file	19.94 MB	$\approx$ 7.53 GB	Phoenix $\approx 378\times$ smaller

Table 6: Preliminary owner-run server comparison; exploratory evidence tier. Raw predictions, sample counts, the preprocessing manifest, and the WER and exact-line denominators were not retained, so these aggregates are not independently reproducible. The four-domain macro CER is an unweighted mean across Agapet, Muharaf, Omar, and RASAM and includes Omar, which is task-matched to Baseer__Nakba through the NAKBA shared task [ 5 ] . A descriptive macro over the three remaining domains is reported alongside it. Baseer attains the lower Omar CER and the higher exact-line accuracy. These values must not be combined with Table 3 or Table 4 .

Figure 4: Oracle@K headroom curve evaluated on the full 1,500-line candidate pool across five domains. Candidate headroom expands with candidate depth but shows diminishing returns with increasing

K

Method	Dev CER (%)	Scientific Interpretation
Greedy	8.131	Baseline local path selection
Beam (width 10)	7.966	Standard CTC beam search
Length-only control	7.977	Length normalization alone does not explain gains
Neural Reranker v1	7.883	BiGRU with listwise soft-target training (small dev gain, $-0.083\,\text{pp}{}$ )
Neural Reranker v2	7.901	Hybrid listwise + pairwise loss (does not outperform v1)
Medoid consensus	8.007	Hypothesis-space consensus harms top-1 selection

Table 7: Consolidated decoder, neural reranker, and quality estimator benchmark on the frozen 250-line development diagnostic (the exact producer checkpoint could not be recovered from the retained historical logs; Oracle@25 is an unachievable diagnostic upper bound).

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Configuration	RASAM-test CER (%)	TariMa-test CER (%)
Greedy (no LM)	7.48	10.27
Order-8 LM	6.76	8.68
$\Delta$	$-0.72\,\text{pp}{}$	$-1.59\,\text{pp}{}$

Appendix

Table 8: Historical exp6 character

n

-gram LM benchmark.

Decoder Mode	D1 Clean	D2 Demo	D3 Auto	Mean CER (%)
Greedy (no LM)	0.39	4.34	3.61	2.78
Always-LM	0.56	2.56	2.28	1.80
Gated LM (conf $<0.95$ )	0.30	3.29	3.41	2.33

Appendix

Table 9: Confidence-gating safety probe across diagnostic sets.

Top Percentile of Lines	Share of Recoverable Errors (%)
Top 5%	12.09
Top 10%	21.63
Top 20%	37.61
Top 30%	53.60
Top 50%	76.24

Appendix

Table 10: Cumulative concentration of recoverable oracle errors on the 250-line development set.

Domain	Beam CER (%)	Oracle@25 CER (%)	Gap ( $\Delta$ pp)
Agapet	8.923	6.223	2.700
Muharaf	12.232	9.904	2.328
Omar	7.858	5.280	2.578
RASAM	7.424	5.455	1.969
TariMa	2.862	1.376	1.486
Unweighted Mean	7.860	5.648	2.212

Appendix

Table 11: Oracle@25 headroom broken down by manuscript domain across the full 1,500-line candidate pool (300 lines per domain).

Domain	Validation NLL	Validation PPL
Agapet	2.472	11.85
Muharaf	2.640	14.02
Omar	2.700	14.87
RASAM	2.520	12.43
TariMa	2.506	12.26
Overall	2.568	13.04

Appendix

Table 12: Per-domain validation perplexities for the neural character LM.

Historical Document Processing

Historical document processing uses AI to transcribe, structure, restore, and analyze manuscripts, books, newspapers, and archival records.

See all 76 papers

Does a Modern-Handwriting Warm-Up Help Historical Arabic OCR? A Reproducible, Compute-Matched Evaluation on Muharaf and KHATT
Aug 23, 2026Sumaih Almarshad, Maram Alamri, Dona Aloraini +4Optical Character Recognition Historical Document Processing
Dal Research Team, Riyadh, Saudi Arabia
RefLAM: A Reference-Grounded Line Annotation Pipeline for Historical Arabic Manuscripts
Aug 25, 2026Mohamed Guechaoui, Mohamed Diaa Zellagui, Souleyman Chaib +1Historical Document Analysis Human-in-the-Loop Annotation
Higher School of Computer Science (ESI-SBA), Sidi Bel Abbes, Algeria
AraMS-28k: The Largest Publicly Released Line-Level Dataset of Historical Arabic Manuscripts with Margin and Insertion-Anchor Annotations
Aug 27, 2026Mohamed Guechaoui, Mohamed Diaa Zellagui, Souleyman Chaib +1Historical Document Processing Optical Character Recognition
Higher School of Computer Science (ESI-SBA), Sidi Bel Abbes, Algeria

Aug 23, 2026J/K move · Enter open · S save

Does a Modern-Handwriting Warm-Up Help Historical Arabic OCR? A Reproducible, Compute-Matched Evaluation on Muharaf and KHATT

Sumaih Almarshad, Maram Alamri, Dona Aloraini, Fares Altuwaim+3

Dal Research Team, Riyadh, Saudi Arabia

Optical Character Recognition Historical Document Processing Cross-Domain Transfer Learning

Open paper AskPDF arXiv

Whether an intermediate stage of modern Arabic handwriting helps or hurts historical Arabic HTR is usually decided from one implementation and one comparison, too thin a basis for a claim either way. We test stability by running the same nominal ablation four times, letting the base checkpoint, encoder-freezing strategy, epoch budget, precision, and learning-rate schedule vary as they naturally did during development, while holding the normalization, scorer, and interval estimation fixed. Each run compares intermediate training on modern handwriting (KHATT) then fine-tuning on historical manuscripts (Muharaf) against fine-tuning on Muharaf directly. Across the four runs the estimated effect swings from -17.64 to +14.52 CER points and reverses sign. The two extremes are exactly the two runs with an identifiable confound (a fivefold lower learning rate in one; a checkpoint of undisclosed provenance in the other); the two clean runs land at -0.25 and +0.94, i.e. no effect. A tight interval from one implementation says nothing about the next. We then run a compute-matched experiment with identical budgets over three seeds: KHATT warm-up is +2.42 CER points worse than a matched same-domain control (95% interval [+0.60, +4.25]); the part of that gap specific to the handwriting domain is only about 0.6 points a small negative effect under this configuration, not a universal result. We release a SaudiHeritage-OCR package with the normalizer, interval scorer, a verified KHATT decoder, experimental manifests, VLM baselines, and an edition-alignment protocol, so the result can be checked independently. The Al-Mahd inscription line is held strictly out and is not offered as a benchmark.

Abstract

Historical Document Processing

Explore similar work

Does a Modern-Handwriting Warm-Up Help Historical Arabic OCR? A Reproducible, Compute-Matched Evaluation on Muharaf and KHATT

RefLAM: A Reference-Grounded Line Annotation Pipeline for Historical Arabic Manuscripts

AraMS-28k: The Largest Publicly Released Line-Level Dataset of Historical Arabic Manuscripts with Margin and Insertion-Anchor Annotations