Navigating the Digitization Gap: An Indirect Evidence Synthesis of AI Methods for Low-Resource Chagatai Manuscripts

Abstract

Many historical handwritten records in low-resource languages remain difficult to access through modern digital systems. This limits efforts to preserve and study cultural heritage at scale. Chagatai manuscripts exemplify these challenges within the Eastern Turki tradition. For centuries, it served as a major written language across Central Asia and supported a rich literary tradition. Large collections of Chagatai manuscripts still survive today, yet only a small amount of this material exists in digital form. As the technical literature specifically focused on Chagatai-HTR remains in its nascent stage, this review synthesizes indirect evidence from taxonomically related Perso-Arabic scripts to establish a foundational research framework. This article presents a systematic literature review following the PRISMA guidelines to examine artificial intelligence methods for handwritten text recognition (HTR) and text restoration in low-resource languages. Analyzing 50 studies published between 2020 and 2026, the review categorizes research trends into handwritten text recognition (HTR), optical character recognition (OCR), script classification, dataset development, and multimodal vision–language systems. The findings reveal a significant architectural shift from traditional segmentation-based CNN and RNN models toward transformer architectures and multimodal approaches. However, for Chagatai specifically, the primary obstacle is not the lack of advanced models but a critical scarcity of basic research infrastructure, including expert-verified transcriptions, annotation standards, and open benchmark datasets. Consequently, this article proposes a concrete development roadmap focusing on systematic digitization, expert annotation, transfer learning, and the creation of baseline models to enable reproducible evaluations.

1. Introduction

Historical manuscripts do more than preserve text; they document cultural identity, language development, and shared memory. Preserving these works increasingly relies on digital repositories. These platforms provide long-term storage, remote access, and opportunities for computational analysis. Digitization alone does not make manuscripts fully accessible. Many historical documents remain difficult to read automatically, search, transcribe, or analyze computationally. It should be noted that while Chagatai is a culturally significant tradition, direct computational research is currently limited. Consequently, this manuscript frames the synthesis of existing methods from neighboring low-resource domains as a strategic roadmap for future Chagatai manuscript digitization, rather than a survey of a mature Chagatai-specific research body.

Global cultural heritage now depends on digital archives. This transition represents a significant advancement in digital preservation paradigms [1]. Optical Character Recognition (OCR) performs well on printed texts in major languages. Handwritten historical manuscripts still resist accurate transcription [2,3]. The gap widens with linguistic complexity and document damage. It is strongest for low-resource languages and scripts that differ from modern digital standards [4,5].

Chagatai, a Turkic literary language used across Central Asia from the 14th to 20th centuries, preserves a substantial body of historical, scientific, religious, administrative, and literary texts [6]. Scholars usually divide its development into several historical periods that shaped its role as one of the main written languages of the region. Similar to Ottoman Turkish, Chagatai has a rich manuscript tradition; however, a large part of this heritage remains outside modern computational analysis. Central to this tradition is the Gunnar Jarring Collection at Lund University, which represents one of the world’s most significant repositories of late Chagatai and Eastern Turki manuscripts. These records are vital for understanding the 18th to 20th century linguistic landscape of Central Asia and present significant challenges for digitization due to severe ink bleed-through and orthographic variability, which are directly addressed in the methodological framework of this study. Transforming these materials into searchable and machine-readable digital collections is a necessary prerequisite for recovering regional knowledge and reintegrating Central Asian intellectual traditions into the global digital record [7,8].

Despite rapid progress in deep learning, a significant gap remains between high-resource languages, such as English, French, and Spanish, and historical or minority languages with limited annotated data [9]. The main reason is data scarcity. Modern sequence-to-sequence models, convolutional-recurrent architectures, and Vision Transformer-based approaches require sufficiently large and representative datasets for effective training and evaluation [10]. Well-known datasets such as IAM and RIMES provide extensive samples for Latin-based handwriting recognition, whereas comparable corpora for Chagatai and related historical Turkic scripts remain scarce, fragmented, and insufficiently standardized. In addition, few pretrained models are specifically designed for ancient or historical scripts. As a result, researchers often adapt models trained on modern Arabic or Persian data, although such models may fail to capture important palaeographic features, orthographic variability, and morphological patterns of historical Turkic texts [11].

Optical Character Recognition (OCR) has achieved high accuracy for machine-printed texts under controlled conditions. However, Handwritten Text Recognition (HTR) remains a more complex task because it must process the natural variability of human writing [12]. While conventional OCR systems are often designed for clean documents with relatively fixed layouts and standardized character forms, HTR must recognize variation in character shape, position, spacing, stroke thickness, and writing style [13]. Since handwriting lacks the regularity of printed text, standard OCR techniques often perform poorly on irregular, degraded, or stylistically diverse manuscript materials [12,13].

The recognition of Chagatai manuscripts is further complicated by the Arabic-based writing system used in many historical sources. Such scripts are cursive, context-dependent, and visually dense [12,13]. A single letter may take different forms depending on its position in a word: isolated, initial, medial, or final. This strong cursiveness creates segmentation ambiguity and may cause vertical overlap between neighboring characters, words, or diacritical marks [13]. Dots and diacritics add another layer of complexity, since a limited set of base letter shapes can represent a larger set of phonemic distinctions. Moreover, variation in calligraphic styles, including Naskh and Thuluth, increases visual diversity and reduces recognition accuracy [12].

Historical manuscript digitization also faces challenges beyond writing style variation [12,14,15]. Aging, environmental exposure, and physical damage often produce stains, faded ink, cracks, paper deformation, and ink bleed-through from the reverse side of a page [15]. These effects obscure baselines, blur character contours, and reduce the contrast between text and background [14]. Therefore, recognition alone is not sufficient. Restoration and preprocessing become integral parts of the overall task. HTR systems for historical manuscripts usually require adaptive binarization, noise removal, skew correction, contrast enhancement, and background normalization [14,15]. These steps help separate text from degraded backgrounds and improve the long-term preservation and computational accessibility of handwritten cultural materials.

This study presents a systematic literature review of artificial intelligence methods for the recognition and restoration of handwritten texts in low-resource languages, with a specific focus on the Chagatai script. The main contributions of the study are as follows:

• A comprehensive taxonomy of current HTR and manuscript restoration methods is proposed, including segmentation-based approaches, neural sequence models, transformer-based architectures, and image restoration techniques.
• Low-resource learning strategies are analyzed, including transfer learning, data augmentation, synthetic data generation, few-shot learning, and human-in-the-loop annotation.
• Reported evaluation metrics, including Character Error Rate (CER), Word Error Rate (WER), recognition accuracy, and restoration quality indicators, are summarized across relevant studies on Arabic-based and Perso-Arabic Turkic scripts.
• Key research gaps are identified, including the lack of standardized Chagatai datasets, limited pretrained models, insufficient benchmark protocols, and the absence of integrated recognition–restoration pipelines.
• A strategic research roadmap is proposed for future work, emphasizing open-source corpora, expert-assisted annotation, few-shot transformer models, and integrated AI pipelines for manuscript restoration and recognition.

Evidence indicates that the recognition of Chagatai manuscripts is constrained by a multi-dimensional intersection of linguistic, visual, and paleographic variables. These include the Arabic-based cursive writing system, positional letter forms, diacritics, orthographic variation, physical manuscript degradation, limited annotated data, and the absence of standardized evaluation benchmarks. The main challenges that define Chagatai HTR are summarized in Figure 1.

These challenges indicate that Chagatai HTR cannot be addressed only as a conventional OCR task. It requires a broader analysis of recognition methods, restoration techniques, data creation strategies, evaluation protocols, and adaptation approaches for low-resource historical scripts. To address this scholarly fragmentation, this SLR systematically investigates: (RQ1) the efficacy of current HTR/restoration paradigms in historical contexts; (RQ2) the availability of standardized benchmarks for Perso-Arabic scripts; (RQ3) script-specific methodological barriers for Chagatai; and (RQ4) the requirements for a reproducible research roadmap.

• RQ1. What artificial intelligence methods are currently used for handwritten text recognition and restoration in low-resource and historical scripts?
• RQ2. What datasets, benchmarks, and evaluation metrics are used to assess HTR systems for Arabic-based and Perso-Arabic Turkic scripts?
• RQ3. What methodological limitations remain in the recognition and restoration of Chagatai handwritten manuscripts?
• RQ4. What future research directions can support the development of reliable AI-based HTR systems for Chagatai and related low-resource scripts?

The remainder of this article is organized as follows: Section 2 details the review methodology, including the PRISMA-based selection process, search strategy, and eligibility criteria. Section 3 presents the core results and taxonomy of AI methods for HTR, character classification, and text restoration, while also synthesizing the available datasets and benchmarking metrics for low-resource scripts. Section 4 presents a critical discussion of the findings, highlights major research gaps, and outlines a research roadmap for Chagatai HTR. Section 5 discusses the study’s limitations. Section 6 presents the final conclusions.

2. Materials and Methods

This review examined recent artificial intelligence methods used for manuscript recognition in low-resource languages. The study followed the PRISMA 2020 framework and applied a structured procedure for literature identification, screening, and selection; the completed PRISMA 2020 checklist is provided in Supplementary Materials File S1. A total of 50 peer-reviewed articles and preprints published from January 2020 to March 2026 were included. This time frame captured recent advances in multimodal architectures and vision–language models. Relevant records were collected from major academic databases and indexing platforms, including IEEE Xplore, EBSCO, and Google Scholar. The search process relied on a predefined Boolean query designed to identify studies related to manuscript digitization, handwritten text recognition, and artificial intelligence techniques for low-resource languages. The resulting body of evidence forms the basis for a reproducible research agenda in this field. The next subsection presents the main objectives of the review and defines the research questions that guide the analysis.

2.1. Study Objectives and Research Questions

The primary objective of this systematic review is to synthesize the current state of artificial intelligence for the recognition and restoration of handwritten texts in low-resource contexts, with an emphasis on identifying transferable methods for the Chagatai script. This study identifies existing research gaps, methodological limitations, and promising directions for building a sustainable research infrastructure. To fulfill this aim, the review addresses the four specific research questions (RQ1–RQ4) established at the end of the Introduction. These questions serve as the foundational framework for the search strategy (Section 2.3) and the eligibility criteria (Section 2.4) used to select and analyze the 50 studies included in this synthesis.

2.2. PRISMA

This systematic review followed the PRISMA 2020 guidelines. The review process used these guidelines to support clear reporting, consistent methods, and reproducible results [16]. All structural stages of the literature workflow, encompassing identification, screening, eligibility evaluation, and final inclusion, adhered strictly to these established reporting standards. Due to the specific nature of this study, which focuses on secondary data analysis of existing peer-reviewed academic literature within computer science and digital humanities without clinical or health-related outcomes, formal registration in international registries such as PROSPERO was not applicable. The complete screening and extraction workflow is visually documented using the standard PRISMA flow diagram (Figure 2), which tracks the precise number of records identified, screened, excluded, and ultimately selected for qualitative synthesis.

2.3. Data Sources and Search Strategy

This review focused on publications related to handwritten text recognition (HTR), optical character recognition (OCR), document restoration, script classification, and AI methods for historical and low-resource scripts. A structured search strategy was prepared before the review began.

The search strategy utilized a tiered Boolean approach across major databases (IEEE Xplore, Scopus, Google Scholar, EBSCO, Web of Science) between 1 January 2020, and 1 March 2026. Search strings were tailored to the specific syntax of each database to ensure maximum retrieval and reproducibility. For Scopus, the query utilized the comprehensive ALL field: ALL ((“handwritten text recognition” OR “handwritten OCR” OR “HTR”) AND (“generative AI” OR “transformer” OR “multimodal”) AND (“low-resource language” OR “Chagatai” OR “historic manuscript”) AND (“restoration” OR “preservation”)). The IEEE Xplore strategy applied “Full Text & Metadata” filters across all Boolean groups, while Google Scholar and EBSCO followed a standard grouped Boolean template (see Figure 3). To ensure transparency and minimize selection bias, the screening process was conducted independently by two reviewers (Z.B. and D.R.). A high level of inter-rater agreement was maintained through this independent double-screening and formal arbitration. In cases where disagreements regarding study eligibility occurred, a formal consensus-building procedure was followed, with a third author (S.B.) acting as an arbitrator to reach a final decision based on the predefined inclusion and exclusion criteria. Data extraction was performed using a standardized coding framework to ensure consistency, evaluating parameters such as model architecture, dataset characteristics, and evaluation metrics. To ensure the reliability of the synthesized evidence and address the inherent heterogeneity of the corpus, each study was subjected to a formal Evidence Tiering approach and structured quality appraisal. Peer-reviewed articles were evaluated based on their methodological advancement, while the inclusion of preprints from repositories (e.g., arXiv, SSRN, TechRxiv) was a deliberate choice to capture high-velocity AI advancements, such as MLLMs, that have yet to complete the formal peer-review cycle. The status and quality tier of all 50 included studies, categorized according to methodological rigor, data transparency, and technical reproducibility, are documented in

Table A1. Mapping of included primary studies to bibliographic references (N = 50).

ID	Year	Title	Source	Ref.	Status
S1	2025	Towards Accurate Recognition of Historical Arabic Manuscripts: A Novel Dataset and a Generalizable Pipeline	ACM Transactions on Asian and Low-Resource Language Information Processing	[22]	Peer-reviewed
S2	2025	Handwritten Text Recognition for Low Resource Languages	ArXiv	[25]	Preprint
S3	2025	Digitizing Nepal’s Written Heritage: A Comprehensive HTR Pipeline for Old Nepali Manuscripts	ArXiv	[26]	Preprint
S4	2024	A Transformer-Based Approach for Arabic Offline Handwritten Text Recognition	Signal, Image and Video Processing	[27]	Peer-reviewed
S5	2024	ET-Network: A Novel Efficient Transformer Deep Learning Model for Automated Urdu Handwritten Text Recognition	PLOS ONE	[28]	Peer-reviewed
S6	2024	Hatformer: Historic Handwritten Arabic Text Recognition with Transformers	ArXiv	[17]	Preprint
S7	2025	CHURRO: Making History Readable with an Open-Weight Large Vision-Language Model for High-Accuracy, Low-Cost Historical Text Recognition	Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing	[29]	Peer-reviewed
S8	2025	TDA-ViT: A Transformer-Based Framework for Unified Urdu Text Recognition via Topological and Visual Feature Fusion	IEEE Access	[30]	Peer-reviewed
S9	2024	Qalam: A Multimodal LLM for Arabic Optical Character and Handwriting Recognition	Proceedings of the Second Arabic Natural Language Processing Conference	[21]	Peer-reviewed
S10	2023	Computationally Efficient Recognition of Unconstrained Handwritten Urdu Script Using BERT with Vision Transformers	Neural Computing and Applications	[31]	Peer-reviewed
S11	2026	SCRAT-Net: An Attention-Based Deep Learning Architecture for Handwritten Word Recognition	IEEE Access	[32]	Peer-reviewed
S12	2024	Machine Learning Approach for Arabic Handwritten Recognition	Applied Sciences	[33]	Peer-reviewed
S13	2025	Bilingual Bangla OCR for Rural Empowerment: Detecting Handwritten Queries and Agricultural Assistance	IEEE Open Journal of the Computer Society	[34]	Peer-reviewed
S14	2024	ADOCRNet: A Deep Learning OCR for Arabic Documents Recognition	IEEE Access	[35]	Peer-reviewed
S15	2024	Optical character recognition for Telugu handwritten text using SqueezeNet convolutional neural networks model	International Journal of Advances in Applied Sciences	[36]	Peer-reviewed
S16	2023	Improving OCR Accuracy for Kazakh Handwriting Recognition Using GAN Models	Applied Sciences	[37]	Peer-reviewed
S17	2025	MSBNet: Handwritten Bangla Character Recognition Using Lightweight Multi-scale CNN Architecture	Proceedings of the 8th CODS-COMAD	[41]	Peer-reviewed
S18	2021	A Fusion-Based Hybrid-Feature Approach for Recognition of Unconstrained Offline Handwritten Hindi Characters	Future Internet	[42]	Peer-reviewed
S19	2026	TifinNet: CNN–Transformer Hybrid Architecture for Tifinagh Handwritten Recognition	IEEE Access	[43]	Peer-reviewed
S20	2025	Multi-Attention Based Convolutional Neural Network for Tamil Handwritten Character Recognition	IEEE Access	[44]	Peer-reviewed
S21	2026	Deep Inception Neural Network with Residual Connections for Tamil Handwritten Character Recognition	Scientific Reports	[45]	Peer-reviewed
S22	2025	A Two-Level Multi-Branch Convolutional Neural Network Framework for Handwritten Gujarati Character Recognition	IEEE Access	[46]	Peer-reviewed
S23	2025	Hybrid Model for Classifying Indo-Aryan and Tamil Texts from Historic Manuscripts	Scientific Reports	[47]	Peer-reviewed
S24	2025	Recognition of Handwritten Characters in Birch-Bark Manuscripts via Object Detection	IEEE Access	[48]	Peer-reviewed
S25	2025	Evaluation of Egyptian Hieroglyph Classification Across Diverse Writing Styles	Proceedings of the 33rd ACM International Conference on Multimedia	[49]	Peer-reviewed
S26	2025	A Novel Framework for Saraiki Script Recognition Using Advanced Machine Learning Models (YOLOv8 and CNN)	IEEE Access	[50]	Peer-reviewed
S27	2026	ResTANet: A Deep Residual Neural Architecture for Tamil Handwritten Character Recognition	IEEE Access	[51]	Peer-reviewed
S28	2025	TSTN: Tibetan Spatiotemporal Online Handwritten Character Recognition with Data Augmentation	TechRxiv	[52]	Preprint
S29	2026	Enhanced Multi-Head Attention Mechanism and BGRU for Lexicon-Free Online Arabic Handwriting Recognition	TechRxiv	[23]	Preprint
S30	2023	Online Mongolian Handwriting Recognition Based on Encoder–Decoder Structure with Language Model	Electronics	[53]	Peer-reviewed
S31	2025	Acceleration of Urdu Optical Character Recognition on Zynq UltraScale+ MPSoC Using Deep Convolutional Neural Network	IEEE Access	[54]	Peer-reviewed
S32	2023	A Vietnamese Handwritten Text Recognition Pipeline for Tetanus Medical Records	ICIS 2023 Proceedings	[55]	Peer-reviewed
S33	2024	Stroke-Based Data Augmentation for Enhancing Optical Character Recognition of Ancient Handwritten Scripts	IEEE Access	[56]	Peer-reviewed
S34	2025	Multilingual Handwritten Text Recognition (HTR) Models for Large-scale Processing of Archival Documents in Low-Resourced Arabic-Script Languages	SSRN	[18]	Preprint
S35	2025	The OCR Quest for Generalization: Learning to Recognize Low-Resource Alphabets with Model Editing	ArXiv	[57]	Preprint
S36	2023	Is It an I or an L: Test-Time Adaptation of Text Line Recognition Models	ArXiv	[58]	Preprint
S37	2024	Restoring Ancient Ideograph: A Multimodal Multitask Neural Network Approach	Proceedings of the LREC-COLING 2024	[59]	Peer-reviewed
S38	2021	Handwritten Kazakh and Russian (HKR) database for text recognition	Multimedia Tools and Applications	[61]	Peer-reviewed
S39	2025	TibHCR: A Large-Scale Dataset for Offline Handwritten Tibetan Character Recognition	IEICE Transactions on Information and Systems	[38]	Peer-reviewed
S40	2026	HHCS: A Dataset of Cipher Symbol Annotations From Handwritten Historical Encrypted Documents for Machine Learning Tasks	IEEE Access	[39]	Peer-reviewed
S41	2025	Digital Handling Procedure for Digitalizing and Indexing Ancient Manuscripts	IEEE Access	[60]	Peer-reviewed
S42	2025	Reviving an Endangered Script: Optical Character Recognition for Syriac	Stanford CS231n Papers	[63]	Institutional Archive
S43	2024	Ancient Yi Script Handwriting Sample Repository	Scientific Data	[64]	Peer-reviewed
S44	2022	KOHTD: Kazakh Offline Handwritten Text Dataset	Signal Processing: Image Communication	[40]	Peer-reviewed
S45	2023	A New Dataset for Mongolian Online Handwritten Recognition	Scientific Reports	[65]	Peer-reviewed
S46	2026	KazakhOCR: A Synthetic Benchmark for Evaluating Multimodal Models in Low-Resource Kazakh Script OCR	Proceedings of the 2nd Workshop on NLP for Languages Using Arabic Script	[24]	Peer-reviewed
S47	2020	A Benchmark for Unconstrained Online Handwritten Uyghur Word Recognition	International Journal on Document Analysis and Recognition	[66]	Peer-reviewed
S48	2026	MultiModal-ArabicNER: A deep learning framework for semantic knowledge extraction from degraded historical arabic manuscripts	Scientific African	[19]	Peer-reviewed
S49	2026	The “Digital Maktaba LP”: Proposing a Comprehensive Dataset for Arabic Script OCR Title Pages in the Context of Digital Libraries and Religious Archives	Umanistica Digitale	[62]	Peer-reviewed
S50	2026	A Lightweight Architecture for Tamil Character Recognition from Palm-Leaf Manuscripts Using Real-ESRGAN and Tesseract OCR	International Journal of Drug Delivery Technology	[20]	Peer-reviewed

(Mapping) and

Table A2. Critical quality appraisal of the included studies (N = 50).

ID	MR	DT	RP	Status	Tier
S1	3	3	3	Peer-reviewed	High
S2	2	2	2	Preprint	Medium
S3	2	2	2	Preprint	Medium
S4	3	2	3	Peer-reviewed	High
S5	3	3	3	Peer-reviewed	High
S6	3	2	2	Preprint	Medium
S7	3	3	3	Peer-reviewed	High
S8	3	2	3	Peer-reviewed	High
S9	2	3	2	Peer-reviewed	High
S10	3	2	2	Peer-reviewed	High
S11	3	3	3	Peer-reviewed	High
S12	2	2	2	Peer-reviewed	Medium
S13	2	3	2	Peer-reviewed	High
S14	3	2	3	Peer-reviewed	High
S15	2	2	2	Peer-reviewed	Medium
S16	2	3	2	Peer-reviewed	High
S17	2	2	2	Peer-reviewed	Medium
S18	2	2	2	Peer-reviewed	Medium
S19	3	2	3	Peer-reviewed	High
S20	3	2	3	Peer-reviewed	High
S21	3	2	3	Peer-reviewed	High
S22	2	2	2	Peer-reviewed	Medium
S23	3	3	3	Peer-reviewed	High
S24	2	2	2	Peer-reviewed	Medium
S25	3	2	2	Peer-reviewed	High
S26	3	2	3	Peer-reviewed	High
S27	3	2	2	Peer-reviewed	High
S28	2	2	1	Preprint	Low
S29	2	2	2	Preprint	Medium
S30	2	3	2	Peer-reviewed	High
S31	3	2	3	Peer-reviewed	High
S32	2	1	2	Peer-reviewed	Medium
S33	2	2	2	Peer-reviewed	Medium
S34	2	2	2	Preprint	Medium
S35	2	2	2	Preprint	Medium
S36	2	2	1	Preprint	Low
S37	3	3	3	Peer-reviewed	High
S38	3	3	3	Peer-reviewed	High
S39	2	3	3	Peer-reviewed	High
S40	2	3	3	Peer-reviewed	High
S41	2	2	2	Peer-reviewed	Medium
S42	1	2	2	Institutional Archive	Low
S43	2	3	3	Peer-reviewed	High
S44	3	3	3	Peer-reviewed	High
S45	2	3	3	Peer-reviewed	High
S46	2	3	3	Peer-reviewed	High
S47	3	3	3	Peer-reviewed	High
S48	3	2	2	Peer-reviewed	High
S49	2	3	2	Peer-reviewed	High
S50	2	2	2	Peer-reviewed	Medium

Note: Quality criteria are rated 1–3 (1 = Low, 2 = Medium, 3 = High). MR: Methodological Rigor; DT: Data Trustworthiness; RP: Reproducibility. Tier is determined by the aggregate score.

(Quality Appraisal) in Appendix A.

Figure 3 illustrates the full search framework and selection flow. Due to heterogeneous metadata standards, researchers performed manual deduplication and bibliographic verification to ensure record integrity. The screening stage excluded patents, presentation slides, commercial reports, and studies lacking a clear methodological focus on HTR/OCR architectures or low-resource script constraints.

Figure 4 illustrates the distribution of the selected primary studies across their respective publication sources. The largest proportion of research was identified within the IEEE, totaling 15 studies. This was followed by a substantial contribution from Springer Nature (8 studies) and the ACM and ACL portfolios (6 studies). Four publications appeared in MDPI journals. To reflect recent work, eight studies came from preprint servers such as arXiv, TechRxiv, and SSRN. Six more papers came from other sources, including PLOS ONE, IEICE, IJAAS, ICIS, and Stanford University. This spread of sources covers established peer-reviewed methods as well as newer techniques.

Table 1. Summary of task types in historical handwritten text recognition research.

Task Type	Number of Papers
Handwritten Text Recognition (HTR)	17
Script/Character Classification	11
Data Creation, Digitization, and Benchmarking	11
System-level and Online Recognition	5
Learning Paradigms and Adaptation Strategies	4
Text Restoration and Enhancement	5 *

* Includes two studies with a primary restoration focus and three proxy studies analyzing cursive degradation patterns in Perso-Arabic scripts relevant to Chagatai.

provides an overview of the primary task addressed by each of the 50 included studies.

Figure 5 illustrates the distribution of the 50 selected studies by their publication year, showcasing a significant surge in research activity starting in 2023 and peaking in 2025 with 20 papers. This upward trend reflects a rapidly growing academic interest in the topic over recent years, with 2026 data representing only a partial-year count.

The distribution of languages and scripts in

Table 2. Summary of language target types in historical handwritten text recognition research.

Language Category	Number of Papers
Arabic/Perso–Arabic	14
Indic Scripts	12
Others	6
Semitic/Ancient Scripts	5
Kazakh	5
Tibeto–Burman	4
Multilingual/Mixed Datasets	4

shows a strong focus on a few groups within historical handwritten text recognition. Arabic and Perso-Arabic scripts lead with 14 studies, while Indic scripts follow with 12. These groups form the bulk of the reviewed work and reflect high research activity on South Asian and Middle Eastern manuscripts. Other scripts appear less often and in smaller groups. Semitic and ancient systems, such as Syriac and Egyptian hieroglyphs, account for five studies, the same number as those on the Kazakh language.

Multilingual datasets and Tibeto–Burman scripts, including Tibetan and Old Nepali, each appear in four studies. Six more papers cover a mix of East Asian scripts and other low-resource data. This pattern shows uneven coverage. Arabic and Indic scripts attract most of the research, while Tibeto–Burman and ancient writing systems see limited coverage, which marks clear gaps for further study.

2.4. Exclusion and Inclusion Criteria

The review was conducted in line with the PRISMA 2020 framework to support transparent and reproducible study selection. The process included three main stages: identification, screening, and eligibility assessment, carried out through two separate search pathways. Database searches in Google Scholar, IEEE, and EBSCO produced 558 records. Prior to screening, 29 records were removed, including 19 duplicate entries and 10 items that did not meet the article requirement, such as editorials and posters. This step left 529 articles for screening. A review of titles and abstracts led to the exclusion of 398 records that did not match the scope of the study. The remaining 131 articles underwent full-text evaluation. During eligibility assessment, 97 studies were excluded based on the predefined criteria. Among these, 50 focused on high-resource or extensively researched scripts, 30 dealt solely with machine-printed text, and 17 examined higher-level natural language processing tasks without a script recognition component. As a result, 34 studies were retained from the database search stream. A second search pathway identified 16 additional records through backward snowballing and manual screening in Google Scholar. Full texts were collected for 16 records. Each record met the eligibility criteria. These studies were then added to 34 eligible records from database searches. The final set contained 50 studies included in the review.

Table 3. Inclusion and exclusion criteria for study selection.

Inclusion Criteria	Exclusion Criteria
Studies written in English	Studies written in other languages
Focused on handwritten text recognition (HTR), optical character recognition (OCR), text restoration, character/script classification, or historical document analysis	HTR or OCR applied primarily in commercial/operational settings (banks, offices, ID verification, enterprise documents)
Low-resource language studies (e.g., Chagatai, other endangered or historic scripts)	High-resource languages (e.g., English, French, widely studied scripts)
Applied deep learning or machine learning methods	Studies without ML or deep learning methods
Published in peer-reviewed sources (journals, conferences) or established repository/preprint platforms (e.g., arXiv, TechRxiv, SSRN, and institutional academic archives)	Review papers, non-academic literature, dissertations, or retracted papers.
Published between 1 January 2020 and 1 March 2026	Published outside the selected date range
Experimental or methodological studies that evaluate HTR pipelines, models, or datasets	Studies focused exclusively on NLP, ASR, hardware acceleration, or other tasks not directly related to HTR/OCR for handwritten text

outlines the strict inclusion and exclusion criteria established to maintain a precise and relevant scope for the literature selection process. The synthesis specifically targets peer-reviewed, machine learning-driven methodologies published between 2020 and early 2026 that focus on historical document analysis and handwritten text recognition (HTR) for low-resource languages, while systematically filtering out high-resource scripts, commercial applications, and non-ML studies.

3. Results

The systematic synthesis of the 50 included studies reveals a transformative shift toward unified neural architectures, which are analyzed here through the lenses of recognition paradigms, data-centric adaptation, and their specific transferability to the Chagatai domain.

3.1. Evidence Synthesis and Methodological Transferability

To address the critical scarcity of direct historical literature on Chagatai Handwritten Text Recognition (HTR), this study implements an indirect evidence synthesis framework. Table 4 systematically delineates the boundaries between the primary target domain gap and the available proxy evidence derived from taxonomically and calligraphically related scripts, establishing a formal taxonomy for cross-lingual architectural and paleographical knowledge transfer.

The synthesis presented in Table 4 establishes a formal framework for an indirect evidence synthesis, explicitly necessitated by the current absence of dedicated technical literature within the Chagatai target domain. By identifying a high degree of structural isomorphism between Chagatai and its Perso-Arabic proxies, we justify a cross-lingual knowledge transfer that prioritizes models capable of handling continuous baseline alignment and complex ligaturial structures [17,18]. This matrix acts as the theoretical anchor for our subsequent analysis, shifting the research focus from generic pattern matching to targeted architectural interventions that address specific paleographical bottlenecks, such as positional allomorphy and ink bleed-through.

Furthermore, this systematic mapping allows for a calibrated mitigation of the domain shift inherent in low-resource settings. We do not propose a monolithic application of modern vision–language decoders; rather, we emphasize architectures such as HATFormer or multi-scale generative networks that offer interpretable visual grounding [17,19]. By selecting architectures based on their empirical success in related scripts, including specialized frameworks for stroke restoration [20] and AraBERT-integrated decoders for Ottoman Turkish [18], we ensure that any future Chagatai HTR pipeline remains robust to the stochastic noise, severe degradation, and orthographic drift characteristic of the Jarring Collection and broader Central Asian archival heritage [18,20].

Table 4. Evidence synthesis and methodological transferability matrix for Chagatai HTR development.

Script/Language Domain	Evidence Type	Proposed/Core Architectures	Expected Transferability to Chagatai	Key Engineering & Paleographical Bottleneck
Chagatai/Eastern Turki	Target Domain (Primary Gap)	CRNN with CTC; Vision Transformers (ViT)	N/A (Target): Object of the proposed development roadmap.	Absolute absence of curated digital text–line pairs; severe ink bleed-through; unique Jarring Collection variations.
Arabic/Persian Nasta’liq	Proxy Evidence	HATFormer; SwinV2-RoBERTa [17,21]	High: Shared baseline isomorphism and complex vertical stacking.	Positional allomorphy; fluid vertical cursive; floating diacritics (nuqta) alignment [22,23].
Ottoman Turkish/Old Uzbek	Proxy Evidence	TrOCR + AraBERT; Encoder–Decoder [18]	High: Structural vocabulary alignment and morpho-syntactic features.	Discrepancies in orthographic conventions; archaic spelling drift; inconsistent matres lectionis [18].
Degraded Archaic Cursive	Proxy Evidence	Real-ESRGAN; Multi-scale Framework [19,20]	Medium: Effective for generative restoration but requires visual grounding.	Risk of hallucinating visually unsupported strokes; modification of critical diacritics; linguistic-prior dominance [19,24].

Note: Bold text indicates domain categories and transferability levels; italics denote specific paleographical and linguistic terms.

Table 4. Evidence synthesis and methodological transferability matrix for Chagatai HTR development.

Script/Language Domain	Evidence Type	Proposed/Core Architectures	Expected Transferability to Chagatai	Key Engineering & Paleographical Bottleneck
Chagatai/Eastern Turki	Target Domain (Primary Gap)	CRNN with CTC; Vision Transformers (ViT)	N/A (Target): Object of the proposed development roadmap.	Absolute absence of curated digital text–line pairs; severe ink bleed-through; unique Jarring Collection variations.
Arabic/Persian Nasta’liq	Proxy Evidence	HATFormer; SwinV2-RoBERTa [17,21]	High: Shared baseline isomorphism and complex vertical stacking.	Positional allomorphy; fluid vertical cursive; floating diacritics (nuqta) alignment [22,23].
Ottoman Turkish/Old Uzbek	Proxy Evidence	TrOCR + AraBERT; Encoder–Decoder [18]	High: Structural vocabulary alignment and morpho-syntactic features.	Discrepancies in orthographic conventions; archaic spelling drift; inconsistent matres lectionis [18].
Degraded Archaic Cursive	Proxy Evidence	Real-ESRGAN; Multi-scale Framework [19,20]	Medium: Effective for generative restoration but requires visual grounding.	Risk of hallucinating visually unsupported strokes; modification of critical diacritics; linguistic-prior dominance [19,24].

Note: Bold text indicates domain categories and transferability levels; italics denote specific paleographical and linguistic terms.

3.2. Terminology and Conceptual Framework

To ensure methodological clarity throughout this study, we distinguish between key processes in the HTR pipeline.

Table 5. Conceptual differentiation of terminology utilized in this study.

Process Pair	Term	Definition/Scope
Pre-processing	Restoration	Generative or algorithmic recovery of missing, damaged, or obscured visual information (e.g., ink bleed-through, physical degradation).
	Enhancement	Contrast normalization, sharpening, and noise reduction aimed at improving visual clarity for existing features without generating new data.
Output Level	Recognition	The computational process of mapping visual patterns (pixels) to character or word-level digital identifiers (output of the HTR model).
	Transcription	The higher-order semantic process of transcribing mapped identifiers into a coherent linguistic representation (includes normalization and orthographic correction).

Note: Bold text differentiates process domains and target terminologies from their descriptive scopes.

provides the rigorous definitions applied in this research to differentiate restoration from enhancement, and recognition from transcription.

3.3. Handwritten Text Recognition (HTR)

Handwritten Text Recognition (HTR) for historical manuscripts aims to transform unconstrained scribal images into machine-readable digital sequences while navigating the high-entropy visual noise of archival degradation. In the domain of low-resource, Arabic-based scripts like Chagatai, this objective is compounded by right-to-left connectivity, positional letter-form ambiguity, and physical artifacts such as ink bleed-through or severe fading. The foundational technical workflow conventionally employs a baseline pipeline consisting of adaptive preprocessing for background normalization, followed by systematic character or line segmentation and hierarchical feature extraction. Research in manuscript recognition has shifted from handcrafted feature extraction methods to deep neural networks and vision-based models. Earlier systems relied on manually designed representations and separate processing stages. Current end-to-end models learn textual patterns, layout structure, and visual characteristics directly from manuscript images. This design reduces reliance on segmentation and other intermediate steps. It also limits the accumulation of errors that often occurs in multi-stage recognition pipelines. To effectively navigate these paleographic and degradation-induced complexities, researchers deploy several distinct computational approaches, which are detailed in the following review of recognition paradigms.

3.3.1. Recognition Paradigms

Handwritten text recognition is undergoing a paradigm shift from modular pipelines toward unified end-to-end architectures that directly map complex visual inputs into textual sequences via sequence modeling [22,25,26,27]. This transition reduces dependency on handcrafted feature engineering and replaces modular stages with integrated architectures, reducing errors caused by fragmented processing. Recent methodological directions suggest that this shift enables more robust dependency modeling, allowing systems to internalize linguistic and structural relationships during the learning process rather than through separate lexical post-processing modules [17,28]. Furthermore, the evolution toward vision–language paradigms facilitates segmentation-free frameworks capable of processing unconstrained historical layouts, which significantly mitigates the error compounding typically associated with hierarchical line or word isolation [29,30]. Collectively, these findings indicate that the field is moving toward architectures that leverage attention-based representation to maintain global contextual awareness while navigating the cursive variability and orthographic inconsistency inherent in low-resource archival scripts [21,31].

3.3.2. Representation and Learning Trends

A dominant methodological direction centers on attention-based representation learning to encode fine-grained spatial variability across heterogeneous script structures [31,32]. This is further strengthened through dependency modeling, enabling long-range contextual reasoning over cursive and morphologically flexible writing systems [17,28]. Collectively, these mechanisms constitute a shift toward integrated representation learning capable of resolving structural ambiguity and high intra-class variability in handwritten text [29,30]. Recent methodological extensions suggest that this shift toward segmentation-free frameworks reduces the cumulative errors caused by traditional character isolation [25,26]. Furthermore, the field increasingly leverages multimodal vision–language learning to align high-dimensional visual embeddings with pre-trained linguistic knowledge, providing the necessary inductive bias to navigate the stochastic nature of scribal variations and archaic orthography [21,29]. This integrated approach is increasingly characterized by the fusion of pixel-level precision with structural invariants, such as connectivity and loop formation, ensuring that models remain robust to the local distortions and environmental degradations typical of historical manuscripts [30,33]. This shift toward unified feature spaces is further exemplified by frameworks that integrate degradation-aware visual patterns with hierarchical morphological analysis. For historical Arabic manuscripts, which share structural connectivity constraints with Chagatai, such integrated representation learning allows visual degradation cues (e.g., ink bleeding and stroke fragmentation) to actively inform linguistic decoding, effectively bridging the gap between pixel-level noise and semantic extraction [19]. However, the efficacy of these representations remains sensitive to systemic decoding constraints, where a reliance on closed-vocabulary spaces can introduce a generalization gap that masks the true visual-to-textual understanding of the system [27,32].

3.3.3. Data Creation and Preprocessing Constraints

Binarization, noise removal, and segmentation remain the prerequisite transformations for historical handwritten text recognition [34,35]. These operations create a trade-off: enhancing clarity often removes stroke details essential for recognition, where improved readability often leads to loss of fine-grained stroke-level and calligraphic details critical for recognition robustness [33,36]. Recent field-level evidence indicates that aggressive thresholding frequently proves counterproductive in historical contexts, as the erasure of critical character connectivity markers can degrade sequence accuracy. This suggests a methodological movement toward processing raw visual signals to ensure that morphological nuances remain accessible for integrated architectures [22]. Furthermore, systemic constraints in spatial normalization, such as fixed-resolution resizing without aspect ratio preservation, introduce horizontal compression artifacts that obscure character thickness and increase visual ambiguity [17,30]. Collectively, these findings indicate that stabilizing recognition performance requires a mechanism-level balance where noise is minimized without compromising the topological invariants of complex handwritten scripts.

3.3.4. Synthetic Data and Generalization Limits

Synthetic data is widely adopted to improve scalability by expanding training distributions, yet it fails to generalize effectively to real manuscript conditions due to mismatched degradation patterns and stochastic visual noise [26,37]. This mismatch produces a persistent generalization gap when transitioning from controlled synthetic environments to authentic archival documents [22,29]. This domain gap is further exacerbated by the failure of simulated rendering to model real-world document degradation, such as ink-bleed and paper texture [26,29]. Field-level evidence suggests that while synthetic pre-training provides a necessary inductive bias for visual feature extraction, it often results in higher gains from model personalization rather than script-level generalization, leading to substantial performance degradation when encountering unseen “in-the-wild” historical styles [17,37]. In parallel, lexicon-based or closed-vocabulary decoding introduces systematic decoding bias, inflating apparent performance while restricting generalization to unseen lexical distributions [27,32]. This systemic reliance creates a facade of robustness in low-resource settings, where high recognition rates are frequently an artifact of vocabulary matching rather than true visual-to-textual understanding [32,33]. Ultimately, these constraints indicate that bridging the generalization gap requires moving beyond closed-world supervision toward architectures capable of internalizing the structural and topological invariants of complex handwritten forms [30,34].

3.3.5. Benchmarking, Metrics, and Comparability

Evaluation practices remain fragmented due to heterogeneous benchmarking protocols [21,25]. CER and WER quantify sequence accuracy, mAP captures localization quality, and raw Accuracy often ignores sequence structure entirely, making it insufficient for evaluating complex transcription structure. These metrics evaluate fundamentally different objectives; therefore, results are not directly comparable across studies. Lack of standardized datasets and evaluation protocols leads to limited reproducibility across studies, reinforcing comparability issues [22,26]. This heterogeneity is further amplified when evaluations rely on non-standard or non-public datasets, limiting reproducibility and cross-study validation. Recent methodological trends attempt to mitigate this fragmentation by introducing nuanced indices, such as the Error Stability Index (ESI) and Character Shape Error Rate (CSER), to differentiate between minor diacritical shifts and complete morphological misclassifications in cursive scripts [22,32]. This indicates a systemic effort to move beyond simple edit distance toward metrics that reflect the topological and structural invariants of non-Latin scripts [30]. However, the effectiveness of these evaluation paradigms is often compromised by the limited annotated data available for historical manuscripts, where environmental degradation and script sparsity introduce stochastic noise that current sequence-level metrics fail to distinguish from architectural failure [26,29]. Specifically, high performance in isolated character classification (e.g., >98% Top-1 Accuracy [38,39]) does not imply maturity in line-level HTR, where sequence-level metrics like CER remain significantly higher (e.g., 6.5–38.4% [18,40]) due to cursive connectivity and alignment errors. Generalized claims of “system robustness” are therefore invalid if they mix Accuracy-based classification with CER-based transcription.

3.3.6. Low-Resource and Historical Script Challenges

Historical manuscripts and non-Latin scripts are constrained by the structural issues of sparsity, variability, and environmental degradation under conditions of limited annotated data [17,26]. These conditions necessitate the adaptation of general recognition frameworks to account for the stochastic nature of physical document aging and the calligraphic diversity inherent in heritage documents [29,30]. For the HTR task family, performance degradation in low-resource settings is a direct consequence of the structural mismatch between contemporary sequence modeling distributions and the high-entropy visual characteristics of historical scripts [22,25,27]. To navigate the high visual uncertainty of cursive scripts, current methodological directions often impose decoding constraints that bias outputs toward closed vocabulary spaces, which limits open-world generalization and often masks the true state of visual-to-textual understanding [27,32]. High accuracy in low-resource settings frequently results from vocabulary-matching bias rather than genuine visual-to-textual feature learning [28,31]. Furthermore, technical stabilization efforts are hindered by a signal clarity vs structural fidelity trade-off, where standard machine transformations such as binarization and noise removal frequently result in the irreversible loss of critical morphological markers essential for character differentiation [33,36]. While synthetic data expansion is widely adopted to improve system scalability, it consistently fails to effectively bridge the generalization gap because artificial noise models rarely capture the authentic stochastic nature of archival degradation [26,37]. This disparity between simulated training environments and authentic historical manuscripts ensures that scalability often comes at the cost of realism and downstream robustness [34,35]. Evidence indicates that global HTR generalization may be restricted by a systemic over-reliance on closed-vocabulary decoding frameworks [21,25]. In summary, the recognition of low-resource historical scripts is undergoing a structural shift toward unified, attention-driven architectures that internalize contextual dependencies. However, the domain misalignment between clean synthetic supervision and the diverse physical realities of archival heritages remains an unresolved gap. Ultimately, the system-level limitation of decoding bias persists, where reliance on predetermined vocabulary prevents the realization of truly generalized recognition across written history.

The field is converging toward unified sequence-aware learning paradigms that emphasize contextual dependency modeling over isolated character recognition. However, persistent gaps between synthetic and real manuscript domains, alongside benchmarking fragmentation and decoding bias, remain central limitations to robust generalization in historical and low-resource handwritten text recognition.

Table 6. Comparative analysis of selected studies on HTR.

Study	Representation & Learning Type	Task Scope	Sequence Error Metrics (CER/WER)	Classification/ Localization Metrics (Acc/mAP)	Generalization Constraint
[26]	end-to-end learning	line recognition	CER	Accuracy	limited reproducibility
[22]	end-to-end learning	word recognition	CER, WER	—	domain misalignment
[31]	attention-based representation	line recognition	—	Accuracy	morphological variability
[28]	attention-based representation	line recognition	—	sequence accuracy	structural fidelity loss
[17]	attention-based representation	word recognition	CER	—	morphological variability
[30]	dependency modeling	line recognition	CER, WER	—	fragmented benchmarking
[34]	sequence modeling	word recognition	—	Accuracy, sequence accuracy	structural fidelity loss
[32]	attention-based representation	word recognition	CER, WER	—	closed vocabulary bias
[29]	end-to-end learning	paragraph recognition	—	Accuracy	structural fidelity loss
[27]	dependency modeling	line recognition	CER	—	domain misalignment
[35]	sequence modeling	line recognition	WER	—	fragmented benchmarking
[25]	sequence modeling	paragraph recognition	CER, WER	mAP	fragmented benchmarking
[37]	attention-based representation	word recognition	CER, WER	—	limited reproducibility
[33]	dependency modeling	word recognition	CER	—	morphological variability
[36]	sequence modeling	word recognition	—	Accuracy	fragmented benchmarking
[21]	sequence modeling	line recognition	WER	—	domain misalignment
[19]	integrated representation learning	line recognition	CER, WER	—	archival degradation

provides a comparative overview of recent methodological trends in handwritten text recognition (HTR), categorizing selected studies by their learning approach, task scope, evaluation metrics, and primary generalization constraints. The taxonomy reveals a prevalent shift toward attention-based and sequence modeling techniques predominantly applied to word and line recognition, while highlighting persistent open challenges such as domain misalignment, morphological variability, and fragmented benchmarking.

3.4. Script/Character Classification

Script and character classification within the domain of historical, low-resource manuscripts operates through a dual objective: global script identification to discern broad document typologies and scribal hands, and local character classification aimed at the isolation of individual graphemes and their positional allographs. In the specific context of Arabic-based Chagatai manuscripts, this task is fundamentally complicated by the inherent fluidity of cursive writing and the presence of complex ligatures that frequently defy discrete segmentation boundaries. Researchers face several challenges in historical manuscript analysis. Character shapes often vary widely across scribes, time periods, and writing traditions. At the same time, different characters can appear visually similar. These difficulties become more pronounced in manuscripts affected by fading ink, stains, torn pages, and other forms of physical deterioration. For this reason, character classification remains a key stage in the Handwritten Text Recognition (HTR) process. Classification models can narrow the range of possible interpretations before language-level processing, or support final decoding in systems that require high recognition accuracy. Their performance relies heavily on the quality of the visual representations used to capture and distinguish subtle character features.

3.4.1. Representation Strategies in Script and Character Classification

Script and character classification can be understood as a set of methodological directions shaped by visual granularity, representation strategies, and evaluation practices. A common trend across recent literature indicates a transition from manually engineered heuristics toward sophisticated visual feature abstraction [41,42]. This shift is necessitated by the inherent complexity of non-Latin scripts, where traditional feature extraction often fails to capture the hierarchical dependencies of complex glyphs [42,43]. Representation depth is contingent upon the model’s capacity to resolve hierarchically complex morphological cues.

3.4.2. Structural Similarity and Class Separability

Across studies, the efficacy of pattern discrimination is fundamentally constrained by the structural topology of the script [44,45]. Structural similarity between distinct characters reduces class separability, leading to pervasive classification ambiguity [44,46]. This limitation is particularly acute in scripts characterized by overlapping strokes and recursive geometric primitives, where the boundaries between categories become computationally indistinguishable [43,45].

3.4.3. Preprocessing Constraints

The impact of initial data transformations, such as binarization, noise removal, and segmentation, introduces a critical trade-off between visual clarity and structural distortion [47,48]. While these preprocessing steps are intended to enhance the signal-to-noise ratio, they can inadvertently strip away essential curvature information, thereby reducing the reliability of subsequent categorization [45,48]. This suggests that the integrity of the original stroke is a prerequisite for maintaining stable discrimination boundaries in degraded environments.

3.4.4. Intra-Class Variability and Inter-Class Similarity

A significant challenge identified in contemporary research is the tension between intra-class variability and inter-class similarity [49,50]. Such overlap creates instability in discrimination boundaries, particularly when models encounter a distribution shift from controlled settings to historical or degraded contexts [49]. This causal mechanism leads to a pronounced generalization gap, where the performance of a system collapses when exposed to the diverse stylistic nuances found in real-world handwriting [49,50].

3.4.5. Evaluation Metrics and Reproducibility

Evaluation strategies often fail to account for the skewed nature of linguistic distributions, where class imbalance causes accuracy to be heavily biased toward dominant categories [41,51]. While the F1-score provides a more balanced view, it frequently masks structural confusion between visually related glyphs, whereas precision and recall highlight the asymmetric error trade-offs inherent in misclassification [41,51]. These metrics evaluate fundamentally different objectives; therefore, results are not directly comparable across studies. Furthermore, the lack of standardized datasets and evaluation protocols leads to limited reproducibility across studies, reinforcing comparability issues.

3.4.6. Low-Resource and Historical Script Constraints

In the context of historical manuscripts and non-Latin scripts, systemic constraints further degrade classification reliability. The causal chain of data scarcity and high script diversity in limited annotated data environments creates unstable discrimination, as there is insufficient evidence to model the full range of stylistic fluctuations [47,48,49]. Consequently, the interplay between environmental degradation and linguistic complexity remains the primary barrier to robust identification in low-resource settings. Consequently, Chagatai’s Arabic-based script requires classification models that resolve morphological ambiguity through historical script-specific priors rather than isolated visual features. This shift ensures that recognition is grounded in the structural logic of the script rather than generic pattern matching.

The synthesis of current research reveals that script classification is as much a problem of structural representation as it is of evaluation rigor. The move toward more abstract representations must be matched by protocols that can account for the inherent morphological ambiguity and distribution shifts defining historical and low-resource scripts.

Table 7. Summary of methodological trends in script and character classification research.

Study	Representation Strategy	Classification Focus	Evaluation Metric	Generalization Constraint
[41]	visual feature abstraction	character-level classification	Accuracy + F1-score	limited reproducibility
[49]	intra-class variability modeling	character-level classification	F1-score	generalization gap
[45]	class separability	character-level classification	Accuracy	inter-class similarity interference
[51]	pattern discrimination	multi-granular classification	F1-score + Precision/Recall	intra-class variability instability
[44]	visual feature abstraction	character-level classification	Precision + Recall	data scarcity
[47]	inter-class similarity modeling	multi-granular classification	Accuracy + F1-score + Precision/Recall	fragmented benchmarking
[42]	pattern discrimination	character-level classification	Accuracy + Precision + Recall	class imbalance bias
[43]	intra-class variability modeling	script-level classification	F1-score + Precision + Recall	data scarcity
[50]	intra-class variability modeling	multi-granular classification	Accuracy + Recall	limited reproducibility
[48]	class separability	character-level classification	Accuracy + Precision	fragmented benchmarking
[46]	inter-class similarity modeling	multi-granular classification	F1-score + Recall	inter-class similarity interference

synthesizes the methodological trends in script and character classification research by categorizing contemporary studies according to their representation strategy, classification focus, evaluation metrics, and generalization constraints. The taxonomy demonstrates a primary emphasis on character-level and multi-granular classification utilizing structural modeling strategies, while simultaneously uncovering critical performance bottlenecks such as data scarcity, limited reproducibility, and inter-class similarity interference.

3.5. System-Level and Online Recognition

System-level and online recognition approaches can be grouped into several technical directions: spatial–temporal fusion, sequence-aware recognition, hardware-aware deployment, and multimodal integration. Although no mature system-level pipeline has yet been established for Chagatai manuscript recognition, these studies provide relevant methodological directions for future Chagatai HTR systems, especially in terms of spatial–temporal modeling, sequence-aware decoding, preprocessing, and deployment constraints.

3.5.1. Spatial–Temporal Fusion

A common trend in online handwriting recognition (OHR) is the simultaneous extraction of writing trajectory dynamics and structural layout features. The Tibetan Spatial–Temporal Network (TSTN) utilizes a dual-channel encoder where 1D convolutional layers process stroke trajectories and 2D channels capture overall character shapes, achieving a peak accuracy of 89.26% and a stable average of 85.13% on the MRG-OHTC dataset [52]. In contrast, hybrid architectures for Arabic script integrate bidirectional gated recurrent units (BGRUs) with multi-head attention (MHA) to capture fine-grained local dynamics and global contextual dependencies, reaching a character error rate (CER) of 3.86% on the ADAB dataset [23]. Dual-channel fusion optimizes the trade-off between structural robustness and feature precision [23,52].

3.5.2. Sequence-Aware Recognition

Capturing long-range dependencies is critical for cursive and agglutinative scripts. While traditional recurrent neural networks (RNNs) struggle with vanishing gradients, Transformer-based architectures and sequence-to-sequence (Seq2Seq) models provide a more cohesive solution [23,53]. For Mongolian script, the integration of a pre-trained character-level language model into a Seq2Seq+Attention framework reduced the Word Error Rate (WER) to 20.30% on the MOLHW dataset by enforcing grammatical word formation rules [53]. Similarly, Arabic OHR benefits from MHA mechanisms that outperform purely sequential modeling by capturing complex spatial relationships across entire temporal sequences [23].

3.5.3. Hardware-Aware Deployment

Deployment on the edge necessitates significant optimization to manage the latency–accuracy trade-off. A hardware-accelerated Urdu OCR system on the Xilinx Zynq UltraScale+ MPSoC utilizes INT8 quantization to achieve a 75% model compression with minimal accuracy loss (96.73% FP32 to 94.06% INT8). This system achieves real-time inference of 0.189 ms per character (4886.95 FPS) while consuming only 1.32 W. However, most current research remains software-bound; while software-based models achieve high accuracy, they often exhibit high power consumption and latency unsuitable for resource-constrained environments [54].

3.5.4. Multimodal Integration

Comprehensive system pipelines now integrate HTR with document-level processing. A Vietnamese medical record pipeline (DOCR) combines document-layout analysis, adaptive preprocessing for deblurring, and a hybrid VGG19–Transformer model, yielding a CER of 2% and a WER of 12% [55]. Furthermore, modern Transformer frameworks are beginning to unify online trajectory data and offline image features into a single sequence, as seen in the C-TST and Col-OLHTR models, which map sequential stroke patterns to 2D spatial representations [23,52].

3.5.5. Critical Analysis and System-Level Limitations

A significant research gap exists regarding the standardization of datasets and reporting metrics. While hardware-centric studies prioritize power consumption and throughput [54], algorithmic studies for Arabic, Mongolian, and Vietnamese scripts often omit system-level latency and power metrics entirely [23,53,55]. Furthermore, dataset inconsistency across scripts and the reliance on specialized, private archives such as Vietnamese nursing diaries hinder broad cross-system benchmarking [52,55]. Comparability issues are exacerbated by high computational costs; for instance, hybrid VGG19–Transformer architectures and other attention-based models are reported to be expensive to train due to the quadratic complexity of self-attention on long-duration handwritten trajectories [23,52,55]. Finally, the transition from controlled datasets to real-world clinical environments necessitates elaborate multi-stage preprocessing pipelines to manage blurry or degraded strokes, which introduces additional computational overhead that remains under-reported in current performance evaluations [55].

Table 8. Methodological summary of system-level and online handwriting recognition approaches.

Approach	Core Method	Sequence Error Metrics (CER/WER)	Classification/ Efficiency Metrics	System Benefit	Limitation	Dataset Stand.	Studies
Spatial–Temp- oral Fusion	1D/2D CNN + ResNet	—	89.26% Peak Acc. (Tibetan)	Improved robustness in trajectory features	1D CNNs struggle with complex features	Low; uses script-specific MRG-OHTC	[52]
Sequence-Aware (Hybrid)	BGRU + Multi-Head Attention	3.86% CER (Arabic)	—	Superior long-range dependency capture	High computational cost for attention	Moderate; uses ADAB and KHATT	[23]
Sequence-Aware (LM)	Seq2Seq + Pre-trained LM	20.30% WER (Mongolian)	—	Reduced OOV error rates	Not yet practical for end-to-end apps	Low; uses script-specific MOLHW	[53]
Hardware-Aware	CNN + INT8 Quantization	—	4886.95 FPS; 1.32 W (Urdu)	Real-time edge deployment	2.67% accuracy drop from quantization	High; benchmarks against CPU/GPU	[54]
Multimodal Integration	VGG19 + Transformer	2% CER (Vietnamese)	—	Automated layout and noise handling	Model is expensive to train	Low; private medical records	[55]

summarizes the main system-level and online recognition approaches identified in the reviewed studies. The comparison focuses on the core method, reported performance, system-level benefit, limitation, dataset standardization, and relevance to deployment.

Current research shows that the central challenge is no longer only the distinction between offline and online recognition, but the balance between accuracy, computational efficiency, and deployability. Attention-based models improve contextual understanding, but their computational cost remains a barrier for low-power or real-time HTR systems.

3.6. Learning Paradigms and Adaptation Strategies

Learning paradigms and adaptation strategies in handwritten text recognition can be grouped into several methodological directions. These methodological shifts reflect a transition from fully supervised training on modern high-resource scripts to more flexible, decentralized, and test-time strategies designed to address the challenges of historical manuscripts and low-resource alphabets. The current scarcity of pretrained Chagatai resources suggests that a transition toward cross-lingual transfer and few-shot adaptation from related Perso–Arabic Turkic scripts may be a viable technical pathway [18]. These strategies provide a technically viable pathway to overcome extreme label scarcity without requiring massive, non-existent corpora.

3.6.1. Supervised and Data-Efficient Learning

Supervised learning in low-resource contexts often relies on data augmentation to mitigate label scarcity. A common trend is the use of stroke-based transformations for ancient scripts such as Vattezhuthu, where a 7–10% accuracy improvement was reported by mimicking natural handwriting variations [56]. In contrast to the line-level preprocessing, including dewarping and masking, used for archival Arabic documents to standardize image inputs for Transformer models [18], character-level stroke augmentation focuses on preserving morphological integrity [56]. However, comparison between these studies remains limited because many approaches rely on script-specific rules. For example, stroke-based augmentation for Vattezhuthu may not be directly applicable to the right-to-left connectivity of Ottoman Turkish scripts [18,56]. This suggests that data-efficient HTR in low-resource settings requires a balance between global image rectification and script-specific structural preservation.

3.6.2. Transfer Learning and Pretraining

Transfer learning typically involves using models pretrained on high-resource Latin datasets, such as HierText or IAM, as feature baselines for non-Latin or historical scripts [57,58]. Ref. [18] fine-tuned TrOCR models, originally pretrained on English, for archival Arabic scripts by replacing the default decoder with AraBERT. While [57] used task arithmetic to adapt Latin-pretrained models to historical ciphers, including the Borg and Copiale manuscripts, Ref. [18] emphasized linguistic adaptation for Ottoman Turkish. Data suggest that many computational methods remain influenced by a persistent Latin-centric pretraining bias, which potentially hinders the accurate recognition of non-Western historical documents [18,57]. This suggests that although pretraining provides a useful initial representation, existing models remain constrained by the orthographic patterns of dominant high-resource languages.

3.6.3. Self-Supervised Learning and Test-Time Adaptation

Self-supervised and adaptive paradigms are increasingly applied during inference through test-time adaptation (TTA) to address writer-specific style shifts. Ref. [58] proposed an iterative self-training approach that uses a confidence measure based on local smoothness in the model’s prediction region. In contrast to pseudo-labeling techniques used for historical document generalization [57], TTA uses feedback from a language model (LM) to update only the optical encoder for a single test image [58]. A reported limitation is that excessive iterations of self-training at test time may lead to model divergence by overfitting to inaccurate self-labels, a risk that is not fully evaluated in meta-learning contexts [57,58]. This indicates that balancing local writer context with global language priors is essential for robust recognition of unannotated historical data.

3.6.4. Few-Shot and Low-Resource Adaptation

Few-shot adaptation strategies seek to incorporate unseen alphabets and under-represented scripts by leveraging robust feature representations. Ref. [57] demonstrated that model merging and task arithmetic can support the adaptation of pretrained models to under-represented scripts, although direct comparison with other approaches remains difficult due to differences in datasets and metrics. While [56] addressed low-resource Vattezhuthu script through synthetic data generation, Ref. [57] focused on model editing to incorporate under-represented alphabets such as Bangla and historical ciphers. However, comparability issues arise due to the lack of shared benchmarks for unseen scripts, with studies ranging from ancient stone inscriptions to 18th-century archival registers [18,56,57]. This suggests that decentralized optimization methods, including model merging and task arithmetic, may offer a scalable alternative to centralized pretraining for archival preservation.

3.6.5. Domain Adaptation and Cross-Lingual Transfer

Domain adaptation in archival contexts often involves cross-lingual transfer, where multilingual joint training is used to mitigate data sparsity. Ref. [18] reported that syncretic training with modern Arabic and historical scripts, such as Ottoman Turkish, improves accessibility for cultural heritage research. This is comparable to the adaptation of Latin-pretrained models to the Kazakh language in the KOHTD dataset, where TTA achieved an 8.8% absolute Character Error Rate (CER) improvement [58]. A critical limitation is the linguistic shift between historical and modern variants. For example, Ottoman Turkish may not be understood by modern native speakers without specialized education, even if transcription accuracy is achieved [18]. This suggests that cross-lingual HTR must bridge both orthographic similarity and syntactic evolution.

3.6.6. Visual–Linguistic Integration in Transformer-Based HTR

In modern HTR architectures, sequence modeling typically processes online signals, while the image encoder handles offline inputs [18,58]. This division facilitates end-to-end learning by allowing the Vision Transformer to extract high-dimensional features from static line images, which the language model then decodes by applying linguistic context to the resulting sequences [18,58]. Such frameworks enable joint optimization of optical and linguistic components, often using sequence alignment objectives such as connectionist temporal classification (CTC) or attention-based decoding to align visual representations with textual outputs [57,58].

3.6.7. Critical Analysis and Adaptation Limitations

The current HTR landscape is characterized by dataset inconsistency, with evaluations spanning ancient Vattezhuthu [56], historical ciphers [57], and Kazakh exam papers [58]. There is a notable lack of benchmarks for low-resource adaptation, leading to inconsistent metrics: some studies report accuracy improvements [56], while others emphasize CER reductions [58]. Furthermore, the reliance on Latin-centric pretraining remains a bottleneck, as models may not inherently capture the nuances of right-to-left or highly structured scripts [18,57]. Finally, low-resource validation remains weak, since many models are tested only on small held-out subsets of specific archival collections [18,58].

Table 9. Methodological summary of learning and adaptation strategies for low-resource handwritten text recognition.

Paradigm	Core Strategy	Sequence Error Metrics (CER/WER)	Classification/ Accuracy Metrics	Adaptation Benefit	Limitation	Dataset(s)	Studies	Relevance to Chagatai HTR
Supervised Learning	Stroke-based data augmentation	—	7–10% accuracy improvement compared to baseline	Preserves structural integrity of ancient characters	More complex to design than traditional geometric methods	Vattezhuthu (Ancient South Indian)	[56]	Preserves the calligraphic integrity of Perso-Arabic scripts while artificially expanding sparse datasets through realistic morphological variations.
Test-Time Adaptation	Iterative self-training with LM feedback	16.8% CER (Original); 25.2% CER (Corrupted Avg)	—	Absolute CER improvement of 0.3% to 8.8% for Kazakh	Risk of model divergence if iterations exceed a specific threshold (K = 4)	KOHTD (Kazakh)	[58]	Enables real-time adaptation to unique handwriting signatures of individual historical scribes, which is critical for decoding varied hands in unannotated archival collections.
Meta-Learning	Task arithmetic and model merging	—	Accuracy: 0.591 (Vatican); 0.505 (Borg); 0.462 (Bangla)	Reported to accelerate incorporation of new languages by nearly three times (under specific experimental settings)	Global performance tends to plateau beyond a critical data threshold (approx. 60% of data)	Historical Ciphers (Vatican, Borg, Copiale), MLT19 (Non-Latin)	[57]	Facilitates the rapid incorporation of unseen Chagatai alphabets by merging feature representations from related high-resource models (Arabic/Persian) through task arithmetic.
Multilingual Transfer	Fine-tuned TrOCR with AraBERT joint training	Test CER: 38.4%; Test WER: 49.3%	—	Mitigates data sparsity for Transformer architectures via syncretic training	Syntactic and orthographic shifts prevent understanding by modern native speakers without specialized education	Archival Arabic, Persian, and Ottoman Turkish	[18]	Mitigates data sparsity through syncretic training, leveraging significant linguistic and orthographic similarities between Chagatai, Ottoman Turkish, and Persian.

summarizes the main learning and adaptation strategies identified in the reviewed studies. The comparison highlights the core strategy, reported performance, adaptation benefit, limitation, dataset context, and relevance to low-resource handwritten text recognition.

The current trajectory of HTR research indicates a shift toward decentralized and test-time strategies to overcome the high cost of large-scale annotated data. However, the trade-off between model stability and style-specific adaptation remains a major challenge, especially for historical documents. Significant research gaps remain in the standardization of low-resource benchmarks and the development of pretraining resources that do not rely exclusively on Latin scripts.

3.7. Text Restoration and Enhancement

Text restoration and enhancement in handwritten text recognition (HTR) and OCR can be conceptualized as a group of methods designed to recover degraded, incomplete, or visually corrupted textual structures before or during recognition. Digital document restoration is presented here as an emerging sub-field. The evidence base remains heterogeneous, and this review focuses on potential technical transferability rather than established Chagatai-specific solutions. While direct empirical data for Chagatai restoration is absent, the synthesis of proxy evidence from related Perso–Arabic scripts identifies generative super-resolution and attention-based masking as viable pathways for mitigating the severe ink bleed-through and physical degradation characteristic of Central Asian archives [20,59]. Although Table 1 identifies only two studies with restoration as their primary focus, methods for visual signal enhancement and stroke reconstruction are increasingly integrated as critical components in broader recognition pipelines and multimodal systems. Within the reviewed corpus, Ref. [59] proposes a Multimodal Multitask Restoring Model (MMRM) for ancient character restoration. Although this model demonstrates efficacy in pixel-level reconstruction for isolated ideographs, transferring such character-centric paradigms directly to Perso–Arabic cursive systems is methodologically limited due to the structural requirement to preserve continuous horizontal baselines, mandatory ligatures, and contextual positioning inherent to scripts like Chagatai.

Beyond character masking, generative super-resolution models such as Real-ESRGAN have demonstrated high efficacy in recovering high-frequency details and broken character borders in severely deteriorated manuscript supports [20]. By restoring the structural integrity of faint strokes before OCR processing, these lightweight architectures facilitate significant accuracy boosts on scripts characterized by complex, stylized handwritten forms [20]. This generative approach is particularly vital for historical Perso-Arabic variants where ink fading frequently obscures the fine-grained connectivity markers necessary for accurate decoding.

Recent field-level evidence identifies a shift toward multi-scale and degradation-aware enhancement frameworks specifically tuned for historical Arabic scripts. For instance, integrated frameworks utilizing encoder-decoder architectures with skip connections have been successfully deployed to address faded ink, stains, and bleed-through in manuscripts dating from the 12th to 15th centuries [19]. These models leverage local and global visual context to optimize document enhancement simultaneously with semantic extraction, reducing the performance gap caused by authentic archival degradation [19]. Similarly, for archival Ottoman Turkish and Persian documents, which structurally correspond to the layout complexities of Chagatai, restoration pipelines increasingly rely on strategic line-level dewarping and geometric rectification to straighten warped text lines, combined with median-color masking to smooth background gradients [18].

Furthermore, minimizing error rates in connected text strings requires the integration of spatial attention mechanisms to resolve the “signal clarity vs. structural fidelity” trade-off. The HATFormer architecture utilizes a customized BlockProcessor to standardize text-line heights while preserving aspect ratios, effectively minimizing horizontal information loss that would otherwise obscure the thickness of vertical strokes [17]. By leveraging self-attention and cross-attention maps, these Transformer-based systems can successfully distinguish between connected characters and accurately identify critical diacritical markers (nuqta) by considering their contextual influence on surrounding strokes [17].

Complementing this, architectures like ADOCRNet address morphological variability through geometric data augmentation, including perspective and stretch transformations, which stabilize recognition in unconstrained mixed-font cursive environments [35]. Additionally, for manuscripts with highly irregular glyph contours, polygon-based segmentation and adaptive thresholding using the Sauvola method have proven more reliable than conventional bounding boxes for maintaining stroke clarity and handling irregular text boundaries under severe contrast degradation [60]. To provide a structured comparison of the identified restoration methodologies and their technical relevance to the Eastern Turki tradition,

Table 10. Summary of cursive restoration and enhancement approaches.

Study	Script Target	Core Technical Mechanism	Addressed Degradation	Paleographic Benefit
[17]	Historical Arabic	HATFormer (ViT–RoBERTa) with BlockProcessor and spatial attention maps	horizontal information loss; cursive ambiguity	improves recovery of diacritics and positional character forms
[35]	Arabic documents	ADOCRNet (CNN–BLSTM) with geometric augmentation	morphological variability; font noise	enhances robustness across heterogeneous document styles
[18]	Ottoman Turkish, Persian, and Arabic	TrOCR + AraBERT with line-level dewarping and median masking	curved baselines; color-gradient distortions	restores structural regularity in archival text lines
[60]	Ancient manuscripts (historical proxy)	polygon-based segmentation with adaptive Sauvola thresholding	irregular contours; paper-fiber artifacts	improves delineation of complex glyph boundaries
[20]	Tamil (historical proxy)	Real-ESRGAN (RRDB-based GAN) with adaptive morphological preprocessing	faint/broken strokes; leaf-fiber background noise	recovers high-frequency details and structural integrity of characters

summarizes the core enhancement approaches based on their technical mechanisms and target script characteristics. This synthesis evaluates how diverse neural architectures mitigate specific archival degradations, ranging from horizontal information loss to complex paper-fiber artifacts, and highlights the corresponding paleographic benefits, including improved recovery of positional forms and diacritical integrity, which are critical for robust HTR in degraded historical environments.

3.8. Data Creation, Digitization, and Benchmarking

Data creation, digitization, and benchmarking in handwritten text recognition can be grouped into several methodological directions. These directions include dataset construction, digitization and preprocessing protocols, synthetic data generation, benchmarking practices, and low-resource data infrastructure challenges. For Chagatai manuscript recognition, this issue is especially important because no widely accepted open benchmark, standardized digitization protocol, or sufficiently large annotated corpus has yet been established. As a result, future Chagatai HTR research will require not only model adaptation but also systematic corpus construction, expert-assisted annotation, and reproducible evaluation protocols.

3.8.1. Dataset Creation Strategies

A prevailing trend in low-resource HTR involves a distinction between structured data acquisition through standardized collection forms and the digitization of rare historical artifacts. For scripts with active writer populations, researchers use grid-based sampling or LaTeX-generated forms to support large-scale collection and automated labeling [38,61]. For instance, the use of 10×10 grid sheets in Tibetan HTR allowed manual annotation to be restricted to only 1% of the total samples, with the remaining data extracted automatically [38].

In contrast, historical manuscripts such as ancient Sundanese palm-leaf texts or European encrypted correspondence require high-resolution DSLR photography and expert-driven manual annotation at the glyph level [39,60]. To address the scarcity of specialized datasets for library cataloging, recent frameworks have moved toward hybrid pipelines that leverage Large Vision-Language Models (VLMs). For instance, the “Digital Maktaba LP” project employs models like Qwen-2vl to classify frontispieces and extract enriched metadata from Arabic-script title pages, establishing an initial model-generated baseline that serves as a foundation for iterative expert-led verification and refinement [62]. This shift suggests that while structured sampling can accelerate data creation for living scripts, the digitization of historical cultural heritage remains strongly dependent on specialized human expertise, which can now be augmented by multimodal AI frameworks to handle the visual richness of Visually Rich Documents (VrD). This observation is directly relevant to Chagatai, where manuscript materials are historically valuable, visually heterogeneous, and likely to require expert participation in transcription, annotation, and paleographic interpretation. By leveraging the redundancy of information found in “Frontispiece Pages Groups” (FPG), researchers can utilize normalized scripts from adjacent pages to inform the recognition of more complex, decorative calligraphic styles [62]. Such a multi-layered approach effectively transforms the Arabic-script title page, which is one of the most error-prone components of digital cataloging, into a reliable entry point for bibliographic metadata enrichment and corpus development.

3.8.2. Digitization and Preprocessing Protocols

Preprocessing protocols focus on reducing background interference and normalizing irregular inputs. Adaptive thresholding, particularly the Sauvola method, and binarization are commonly used to isolate text strokes from complex backgrounds such as palm-leaf fibers or degraded paper [60,63]. In cases involving miniature symbols or dense layouts, image slicing techniques such as Slicing Aided Hyper Inference (SAHI) are implemented to improve the visibility of small objects during the detection phase [39].

Additionally, tilt and angle correction are frequently employed to normalize scanned sheet data [38,40,64]. Beyond geometric normalization, researchers have introduced evolutionary strategies, such as Genetic Algorithms (GA), to automate the segmentation of lines and words in unconstrained document layouts [40]. However, the lack of standardized preprocessing protocols across different script environments often limits reproducibility across studies, since models optimized for one denoising or segmentation pipeline may not generalize well to different manuscript conditions.

For Chagatai manuscripts, preprocessing standardization is particularly important because Arabic-based cursive writing is sensitive to stroke connectivity, diacritics, ligatures, and baseline variation. Excessive binarization or segmentation may improve visual clarity but can also remove small script-specific features that are essential for recognition.

3.8.3. Synthetic Data Generation

Synthetic data generation has emerged as an important response to the scarcity of authentic handwritten labels. Existing approaches range from merging character samples into pseudo-sequences [38,65] to generating artificial benchmarks using digital fonts and noise filters to imitate OCR tasks [24]. Augmentation techniques such as SMOTE, Gaussian noise, and random distortions can improve model robustness in certain experimental settings [38,39,40].

However, these strategies have important implications for real-world robustness. Synthetic data may not fully capture the characteristics of real manuscripts, including authentic scribal variation, paper degradation, ink bleed-through, fading, and uneven page texture [24,60]. In addition, closed-lexicon synthetic datasets may produce performance estimates that do not fully reflect the complexity of unconstrained real-world recognition [66]. For Chagatai HTR, synthetic data may be useful as an initial strategy, but it should be combined with real manuscript samples and expert-validated annotations to avoid overestimating model performance.

3.8.4. Benchmarking and Evaluation Practices

The field currently faces comparability challenges caused by differences in task types and evaluation metrics. Results are not directly comparable when detection-based tasks report mean Average Precision (mAP) [39], recognition tasks prioritize Character Error Rate (CER), Word Error Rate (WER), or Sequence Error Rate (SER) [40,61,65], and isolated character classification studies rely on Top-1 Accuracy [38,60].

Recent evaluations suggest that, in some low-resource OCR settings, traditional OCR engines such as Tesseract may still outperform current Multimodal Large Language Models (MLLMs), including Qwen2.5-VL and Llama-3.2-11B, especially when these models are not adapted to the target script [24]. However, such comparisons should be interpreted cautiously because performance may depend on the dataset, script type, preprocessing pipeline, and evaluation protocol.

High performance in isolated character classification may also mask limitations in sequence modeling, particularly for connected or cursive scripts where alignment, context, and segmentation are major sources of error [63,65]. Therefore, future benchmarking for Chagatai should include not only character-level accuracy but also sequence-level metrics such as CER and WER, as well as evaluations under different levels of manuscript degradation.

3.8.5. Low-Resource Data Infrastructure Challenges

The structural complexity of non-Latin scripts requires a technical bridge between input types and model architectures. Sequence models are predominantly adapted for online signals to capture temporal pen-tip coordinates [65,66], whereas image encoders are required for offline inputs to extract features from static manuscript images [63]. Morphological similarities between distinct glyphs, for example in Tibetan or Kazakh, remain a persistent challenge and often require contextual linguistic support or dictionary-based post-processing to reduce misrecognition [38,61].

Establishing a reproducible Chagatai HTR framework depends on a unified infrastructure that couples recognition metrics with expert-verified paleographic data. This move from isolated model testing to integrated infrastructure is the primary prerequisite for field-level progress.

Table 11. Summary of datasets, models, and benchmarking practices in low-resource HTR.

Study	Dataset	Script	Model	Task Type	Sequence Error Metrics (CER/ WER)	Classification/ Localization Metrics (Acc/ mAP/ F1)	Data Strategy
[39]	HHCS	Cipher	YOLOv10	Detection	—	0.984 (mAP50–95, Glyphs)	SMOTE Augmentation
[60]	AMADI	Sundanese	MobileNet V2	Recognition	—	88.5% (Top-1 Accuracy)	Expert-guided labeling
[38]	TibHCR	Tibetan	PP-OCRv4	Recognition	—	99.70% (Accuracy)	Grid sheet design
[64]	Yi Script	Ancient Yi	RegNet	Classification	—	98.1% (Accuracy)	Automatic sampling
[66]	OLHW-UWord	Uyghur	1D-Conv-4	Recognition	83.23% CAR (OOV)	—	Natural handwriting
[65]	MOLHW	Mongolian	Transformer	Recognition	16.969% WER	—	Volunteer application
[61]	HKR	Kazakh	Attention-Gated-CNN-BGRU	Recognition	8.34% CER	—	LaTeX form generation
[24]	Kazakh OCR	Kazakh	Tesseract	Recognition	15.0% CER (Arabic)	—	Fully synthetic
[40]	KOHTD	Kazakh	Flor	Recognition	6.52% CER	—	Exam paper collection
[63]	DASH	Syriac	Transformer	Recognition	—	43.56% (Accuracy)	Scraped manuscript data
[62]	Digital Maktaba LP	Arabic/Persian	Qwen-2vl + Google Vision AI	Page Classification	—	93% (F1-score)	Expert-guided VLM labeling

systematically maps out the experimental landscape of low-resource handwritten text recognition (HTR) by aligning selected studies with their utilized datasets, scripts, models, task types, metrics, and empirical performance. It further evaluates each study’s underlying data strategy and core technical limitations, offering a concise benchmark for assessing how diverse neural architectures cope with data scarcity across historical or endangered scripts.

Current trends in HTR indicate a significant shift toward hybrid data strategies that combine authentic low-resource artifacts with synthetic augmentation to facilitate deep learning convergence. However, a persistent trade-off remains where synthetic data provides efficient training but often fails to achieve high generalization in real-world historical contexts. Benchmarking remains a major challenge for HTR research. Differences in evaluation metrics and script-specific preprocessing methods limit reproducibility and complicate comparisons across studies, particularly for non-Latin writing systems.

For Chagatai manuscripts, future progress requires more than the adaptation of existing HTR models. The field needs a reliable research foundation that includes digitized manuscript collections, expert-validated transcriptions, standardized training and testing datasets, and evaluation procedures that account for both recognition performance and manuscript degradation.

The reviewed evidence indicates that data availability, annotation quality, digitization protocols, and benchmark design are not secondary technical issues but foundational requirements for Chagatai HTR. Figure 6 summarizes the proposed research infrastructure needed to support reproducible recognition and restoration of Chagatai manuscripts.

3.9. Multimodal Learning in HTR and OCR

The core limitation of unimodal Handwritten Text Recognition (HTR) systems lies in their exclusive reliance on either visual features or standalone linguistic models, both of which lack the cross-modal grounding necessary to resolve severe document degradation, paleographic ambiguity, or ink fading. Multimodal learning addresses these structural constraints through the joint optimization of vision–language representations, aligning high-dimensional visual features from the pixel space with semantic embeddings from the linguistic space within a shared latent space to facilitate mutual decoding. For low-resource scripts like Chagatai, this architecture leverages semantic priors to provide a critical inductive bias, effectively guiding the interpretation of ambiguous cursive ligatures and highly damaged graphemes where visual signals remain insufficient. By integrating contextual knowledge directly into the inference process, these systems move beyond the error-compounding limitations of modular pipelines that treat linguistic correction as an auxiliary post-processing stage. This intersection of vision and language appears to have facilitated an architectural shift from separate OCR/HTR engines toward unified, end-to-end vision–language frameworks [17,25,26,59].

3.9.1. Transition Toward Multimodal HTR/OCR Systems

The evolution of Handwritten Text Recognition (HTR) and Optical Character Recognition (OCR) has progressed from modular, segmentation-dependent pipelines toward unified multimodal learning systems that jointly optimize visual and linguistic representations within shared latent spaces. Across the literature, this transition reflects a systematic reduction in post hoc language correction in favor of integrated cross-modal learning, where recognition emerges from joint inference over heterogeneous modalities rather than sequential processing stages [17,21,26,59]. This shift is driven by the need to stabilize pattern discrimination under degraded visual conditions typical of handwritten and historical document data. To mitigate the performance gap in degraded historical corpora, recent frameworks have moved toward degradation-aware multimodal feature fusion. Specifically, the integration of multi-scale document enhancement with hierarchical morphological analysis allows systems to leverage visual degradation patterns to inform textual decoding decisions, significantly improving NER and relation extraction on 12th–15th century Arabic manuscripts [19].

For Chagatai manuscript recognition, multimodal learning is particularly relevant because visual evidence alone may be insufficient under degradation, while language-based correction is difficult due to the absence of large Chagatai language models and standardized corpora. Therefore, future Chagatai HTR systems will likely require carefully balanced visual-linguistic architectures that combine manuscript image evidence, historical orthographic knowledge, and expert-verified textual resources.

3.9.2. Taxonomy of Fusion Strategies

A formal taxonomy of fusion strategies emerges across studies, spanning early fusion, intermediate fusion, late interaction, and hybrid structural fusion. Early fusion approaches, as evidenced in [59], integrate ResNet-based visual representations with RoBERTa embeddings through additive interaction, yielding substantial improvements in ideograph restoration accuracy. However, this form of fusion remains limited in capturing fine-grained inter-modal dependencies. Intermediate fusion dominates contemporary systems, where cross-modal attention mechanisms align visual tokens with linguistic representations dynamically. Refs. [17,25] demonstrate that such alignment reduces ambiguity in cursive and low-resource OCR settings by improving contextual disambiguation of visually similar glyphs. Ref. [47] further show that bilinear feature fusion, specifically Multi-modal Factorized Bilinear fusion, outperforms standard concatenation by modeling multiplicative interdependencies between contextual semantic embeddings and character-level structural patterns, achieving a peak classification accuracy of 98.04%. Building on these intermediate strategies, recent frameworks have introduced adaptive confidence-based fusion to manage the stochastic reliability of visual and textual cues in degraded manuscripts. For instance, Ref. [19] utilizes a token-level gating mechanism that dynamically reweights modal contributions based on local image quality indicators (e.g., edge strength and contrast) and contextual linguistic coherence, ensuring robust decoding even in archival regions where one modality is severely compromised by noise [19]. In contrast, hybrid structural fusion, as in [30], incorporates topological descriptors alongside pixel-level representations, indicating that structural priors enhance robustness under writer variability and deformation-induced noise. For Chagatai, this distinction is important because Arabic-based cursive manuscripts require not only visual–textual fusion but also preservation of structural features such as ligatures, diacritics, baseline variation, and positional letter forms.

3.9.3. Visual-Linguistic Representation Learning

Representation learning exhibits a clear progression from convolutional architectures toward Transformer-based encoder-decoder systems and, more recently, vision–language models. Refs. [21,26] demonstrate that replacing CNN-RNN pipelines with Vision Transformer encoders paired with pretrained language decoders improves transcription robustness in low-resource handwriting scenarios. Ref. [31] further show that ViT-BERT integration enhances ligature recognition by improving global contextual modeling while reducing sequential dependency limitations. This trajectory is extended by [29], where OCR is reframed as end-to-end image-to-text generation using large vision–language models, reducing dependence on explicit segmentation stages. However, Ref. [24] provide counter-evidence that such models may degrade significantly under zero-shot conditions in low-resource scripts, where weak visual-linguistic alignment can lead to linguistically plausible but visually unsupported outputs.

3.9.4. Functional Roles of Language Models

Within multimodal systems, language models assume three functionally distinct roles that reflect increasing levels of integration. In enhancement-based systems, language models reconstruct semantically plausible outputs from degraded visual inputs, as demonstrated in [59]. In decoder-integrated systems, such as those described by [21,25,26] language models impose autoregressive constraints that enforce syntactic coherence over visual feature sequences. In alignment-driven architectures, Ref. [31] show that linguistic supervision directly shapes visual attention distribution, enabling tighter coupling between glyph-level perception and contextual semantics. Across [17,47], these roles are increasingly interdependent, indicating that language modeling is no longer auxiliary but structurally embedded within multimodal inference via bilinear fusion or cross-attention mechanisms. For Chagatai, this role of language models must be approached cautiously. Since Chagatai lacks large-scale standardized corpora and robust pretrained language models, language-based correction may introduce historically or orthographically inaccurate reconstructions if it is not grounded in expert-validated manuscript data.

3.9.5. Empirical Evidence and Performance Constraints

Results from multiple studies show that combining modalities improves recognition performance compared with single-modality baselines. These gains depend on balanced modality input and on how data spreads across sources. Ref. [17] report strong reductions in character error rates with Transformer encoder-decoder systems compared with CNN-RNN baselines. Ref. [25] show better paragraph-level OCR performance in noisy manuscript settings through language-guided refinement of visual embeddings. Ref. [59] further confirm that multimodal fusion improves restoration accuracy under partial visual degradation. However, Ref. [24] highlight that large vision–language models may generate outputs that are linguistically plausible but insufficiently supported by the visual evidence in zero-shot settings. In such cases, linguistic priors may dominate weak visual signals, leading to incorrect transcription. These findings indicate that multimodal gains are not monotonic but depend on stable cross-modal grounding.

3.9.6. Critical Analysis and Implications for Chagatai

A central contradiction across the literature lies in the tension between linguistic regularization and visual fidelity. While stronger language integration improves robustness under noise and ambiguity, it can also induce over-reliance on statistical language priors when visual signals are weak, as evidenced by [24]. Conversely, Ref. [26] demonstrate that in low-resource settings, insufficient annotated data limits the ability of models to learn stable cross-modal mappings, resulting in degraded generalization under domain shift. This establishes a structural constraint: multimodal performance is governed by alignment stability rather than modality addition. A major challenge when applying Large Vision-Language Models to Chagatai HTR is the difficulty of interpreting results due to stochastic hallucinations and linguistic over-regularization. Although multimodal integration is effective for deciphering faded characters, it often leads to language-prior dominance. In this scenario, the model’s internal weights, typically trained on modern high-resource corpora, tend to override authentic historical morphology. For the Eastern Turki tradition, this manifests as a systematic bias toward modern orthographic standards, which may suppress important archaic variations such as non-standardized matres lectionis or dialect-specific suffixation. As a result, high sequence accuracy can sometimes mask a lack of paleographic fidelity, as the model may produce text that is linguistically plausible but historically inaccurate. To address this, future HTR pipelines need to move toward visually grounded decoding, using methods like cross-attention maps to ensure that each generated token is explicitly supported by visual evidence from the manuscript. In synthesis, multimodal HTR and OCR systems converge toward a unified paradigm of joint visual-linguistic inference, where recognition is formulated as integrated representation learning over heterogeneous modalities rather than sequential decoding. Across [17,19,21,24,25,26,29,30,31,47,59], performance improvements are consistently associated with the depth and stability of cross-modal interaction rather than modality presence alone.

Table 12. Summary of multimodal learning approaches in HTR and OCR. Note: A—true multimodal learning with deep fusion or cross-modal representation; B—hybrid feature fusion; C—benchmark or evaluation-only study.

Author (APA)	Modalities Used	Functional System Type	MC 1	Relevance to Chagatai HTR
[21]	Image + Text (ARBERT)	Foundation Model (Vision Encoder–Decoder)	A	Validates a SwinV2-RoBERTa architecture for resolving positional letter-form ambiguity and complex diacritics in Perso-Arabic scripts.
[47]	Semantic (BERT) + Structural Graphemic (LiScribe)	BERT–LiScribe Hybrid Model	B	Utilizes MFB bilinear pooling for script-specific feature extraction, facilitating the critical differentiation between Chagatai and visually similar Perso-Arabic scripts.
[17]	Image (BEiT) + Text (RoBERTa)	Transformer Encoder–Decoder	A	Implements spatial attention in HATFormer to resolve cursive connectivity and context-sensitive character variability inherent in historical manuscripts.
[25]	Image (DeiT) + Text (RoBERTa)	Segmentation-free HTR Model	A	Applies a segmentation-free framework to prevent error compounding in dense Chagatai cursive where manual character isolation is technically unfeasible.
[59]	Image + Text (RoBERTa)	MMRM (Multimodal Restoring Model)	A	Uses multimodal multitask restoration to reconstruct faded characters by fusing visual residues with surrounding linguistic context.
[31]	Image Patches + Text Context	Multi-head Self-attention Model	A	Leverages ligature-level classification via multi-head self-attention, establishing robust recognition units better suited to highly joined Chagatai orthography than character-level models.
[30]	Grayscale Image + Topology (TDA)	Dual-stream Transformer	B	Fuses TDA geometric invariants with raw pixels to ensure cross-writer generalization by capturing stable structural properties across diverse scribal hands.
[24]	Image + Text Prompting	MLLM Benchmarking Framework	C	Identifies misclassification risks in MLLMs for Turkic Arabic-script text, proving the necessity of specialized fine-tuning to prevent language hallucinations.
[26]	Image (Swin) + Text (BERT/GPT-2)	Transformer-based VED Pipeline	A	Adopts a three-stage synthetic-to-authentic transfer pipeline to mitigate extreme data scarcity in low-resource manuscript domains.
[29]	Page Image + Text	Large Vision-Language Model	A	Fine-tunes large vision–language models to enable robust transcription of complex historical document layouts and non-standard reading orders.
[19]	Image + Text (Classical Arabic)	Multi-task Enhancement and Semantic Extraction	A	Employs adaptive confidence-based fusion to leverage visual degradation patterns for informed textual decoding in historical Arabic cursive, providing a robust template for handling authentic archival noise in Chagatai manuscripts.

¹ MC = Multimodal Class.

presents a summary of the multimodal learning methods reviewed in HTR and OCR research. The comparison examines the data modalities, fusion strategies, interactions between visual and language components, system functions, reported evaluation metrics, and levels of multimodal integration. The table distinguishes fully multimodal models from hybrid fusion methods, language-guided recognition systems, and benchmark-focused evaluation studies.

The reviewed literature suggests that multimodal learning represents a robust strategic option for addressing the complexities of Chagatai manuscript recognition [17,21,47]. It combines information from damaged manuscript images with linguistic context from the text. Its success requires Chagatai-specific datasets, expert-checked transcriptions, and models that retain the original visual details of manuscripts. Models should reproduce the source material accurately, not just generate text that appears linguistically correct. Future studies should strengthen the connection between manuscript images, historical spelling patterns, and reliable textual sources. This link can improve both recognition accuracy and manuscript restoration.

4. Discussion

Evidence suggests that while the HTR task family has progressed toward unified sequence-aware architectures [17,30], the restoration task family is increasingly defined by generative super-resolution frameworks focused on pixel-level fidelity [20,59]. Consequently, field-level progress must be disaggregated by task to account for the divergent requirements of sequence-level CER versus visual reconstruction quality.

The architectural evolution identified in this review is synthesized in

Table 13. Comparative analysis of architectural paradigms for low-resource HTR based on qualitative synthesis.

Architecture	Data Efficiency	Generalization	Interpretability	Comp. Complexity	Performance (Harsh)
CNN-RNN	Moderate	Low–Mod.	Moderate	Low	Moderate
Transformers	Low	Mod.–High	Low	Mod.–High	High
ViT	Low	High	Moderate	High	High
VLMs	Very Low	High	Low	Very High	Very High

Note: Indicators are synthesized qualitatively as metrics are non-comparable across task families. ’Performance’ for CNN-RNNs primarily reflects word-level Accuracy in classification tasks [61,64], whereas for Transformers/ViTs, it reflects CER/WER in sequence-to-sequence line recognition [17,27,30]. No direct cross-family metric parity is implied. Bold text in the table body is used exclusively to distinguish the architectural paradigms.

, which provides a comparative analysis of model families across key performance and efficiency dimensions relevant to low-resource historical scripts.

The results also show that CNN-based models remain the most frequent category in the reviewed corpus. This indicates that convolutional architectures are still useful as stable and computationally efficient baselines, especially in low-resource settings. The integration of multimodal vision–language information serves functionally distinct roles across task families:

• HTR and OCR: Multimodal alignment is used to provide linguistic inductive bias, reducing sequence-level CER in unconstrained cursive scripts [21,32].
• Document Restoration: Multimodal fusion aligns visual residues with semantic context to reconstruct missing or faded graphemic structures [19,59].
• Script Classification: Integration targets the resolution of morphological ambiguity between visually similar positional allomorphs via structural descriptors [30,47].

Reviewed studies also show that multimodal learning does not always improve results. Its performance depends on strong alignment across modalities, enough training data, and stable links between visual features and language representations.

Chagatai manuscript recognition depends on these findings. Chagatai is a historical Turkic literary language. It uses an Arabic-based cursive script. Its manuscripts often show damaged pages and faded ink. Spelling varies across texts. Letters change shape depending on position. Ligatures and diacritics appear in many forms. Annotated digital data is limited. Models trained on English, modern Arabic, or Persian often fail on this material. They need changes that fit Chagatai script and its writing patterns.

The review demonstrates that Chagatai handwritten text recognition is not merely a problem of model selection; rather, the system requires a comprehensive research infrastructure. This setup includes digitized manuscript collections and expert checked transcriptions. It also needs clear annotation rules and standard preprocessing steps. Shared benchmark datasets support fair testing across studies. Without these parts, even strong Transformer or multimodal models struggle in practice. Their results become hard to reproduce, compare, or extend to real manuscript data.

Adaptation methods include transfer learning, few-shot learning, cross-lingual transfer, test-time adaptation, and synthetic data generation. These methods may support future Chagatai handwritten text recognition systems. Each method needs careful use. Synthetic data and GAN-generated samples can expand training data. They do not replace real manuscript images. Language models can improve text consistency. They can also produce text that looks correct but lacks support in the image. This issue matters in historical manuscripts. Spelling variation and unclear letters require expert review.

The optimal trajectory for Chagatai HTR is a staged hybrid paradigm, prioritizing data foundations before integrating advanced multimodal architectures. To provide a structured evaluation of the methodological transferability,

Table 14. Chagatai transferability synthesis matrix for identified model families based on proxy evidence.

Approach Family	Script Similarity	Data Requirement	Pre-training Dependency	Degradation Tolerance	Chagatai Suitability
CNN/CNN-RNN	High	High	Low–Medium	Medium	High (Baseline)
ViT (Vision Transformers)	High	High	High	Medium	High (Cursive)
Generative/ Inpainting	N/A (Technical)	Medium	Medium	High	High (Restoration)
VLMs (Vision-Language)	High	Very High	Very High	High	Med–High (Hallucination Risk)

Note: Bold text in the table body is used to highlight the architectural approach families and emphasize their specific suitability designations for Chagatai HTR.

synthesizes the identified model families based on their suitability for the Chagatai script, accounting for dimensions such as script similarity, data requirements, and tolerance to historical degradation.

First, a representative digitized corpus and benchmark should be created. Second, stable baseline models such as CNN or CNN–RNN architectures should be tested. Third, more advanced Transformer, Vision Transformer, and multimodal models should be evaluated under controlled and reproducible conditions. Finally, restoration, synthetic augmentation, and language-assisted recognition should be integrated only after their contribution is validated on real manuscript samples.

Research Roadmap for Chagatai HTR

The review shows that Chagatai handwritten text recognition needs a staged research plan that draws on multiple fields. High resource HTR systems often build on existing datasets and shared benchmarks. Chagatai work does not have this base. It lacks standard corpora, expert checked transcriptions, and shared annotation rules. It also lacks stable evaluation methods. Future work should not focus only on model testing. It should follow a shared plan that connects manuscript studies, data building, machine learning, and evaluation design.Based on the systematic synthesis of indirect evidence from related Perso–Arabic and Turkic domains, this article establishes a theoretically grounded research roadmap for the development of Chagatai HTR infrastructure. This pathway provides a concrete framework for implementation, derived directly from the synthesized methodological trends and identified research gaps.

Stage 1: Manuscript Inventory and Codicological Digitization. The initial phase focuses on the systematic digitization of Chagatai materials, prioritizing an understanding of the physical degradation common in Central Asian archives. Unlike modern documents, Chagatai manuscripts are often written on regional paper types (such as Samarkand rag paper or various localized Central Asian variants) that exhibit distinct aging patterns. A primary technical hurdle at this stage is the severe acidic ink bleed-through and “show-through” typical of historical Islamic manuscripts, where the corrosive nature of the ink causes text from the verso side to merge visually with the recto side. Digitization protocols must therefore move beyond simple resolution standards to capture the high-entropy visual noise of archival degradation, ensuring a stable image base that accounts for paper deformation, complex stains, and faded ink before any recognition is attempted.

Stage 2: Paleographic Annotation and Cursive Transcription Standards. The second stage establishes expert-driven annotation guidelines that address the specific paleographic complexities of the Chagatai Arabic-based script. Expert paleographers must define standards for handling high positional allomorphy, where characters such as ’alif or sīn drastically shift their graphemic structures across isolated, initial, medial, and final positions. Annotation must explicitly account for complex horizontal and vertical ligatures, particularly those involving letters such as kāf, lām, or mīm, which often merge seamlessly into neighboring glyphs and defy discrete segmentation. Furthermore, the guidelines must provide a logic for ambiguous or floating diacritics (nuqta). In styles such as Nasta’liq, which are common in Chagatai literary works, these marks often drift far from their base characters or overlap with neighboring strokes, requiring a transcription hierarchy that distinguishes between scribal omission and physical fading.

Stage 3: Baseline CNN and CNN-RNN Benchmarking via CTC Loss. Before deploying complex architectures, it is necessary to establish stable baselines using Convolutional Neural Networks (CNN) and CNN-Recurrent Neural Network (CNN-RNN) hybrids. In the context of Chagatai paleography, these models serve a specific purpose: mapping the continuous horizontal baseline of the cursive script. By utilizing Connectionist Temporal Classification (CTC) loss, the system can learn to align the visual signal with the textual sequence without the need for individual character segmentation. This “segmentation-free” approach is vital for Chagatai, as it allows the model to internalize the fluid, connected nature of the writing system before more advanced attention-driven models are introduced to resolve fine-grained morphological and ligaturial details.

Stage 4: Cross-Lingual Transfer and Adaptation to Orthographic Drift. Stage 4 addresses the profound linguistic hybridity and orthographic drift inherent in Chagatai. To provide a rigorous theoretical grounding for this stage, future architectures must leverage the evolving landscape of Large Language Model (LLM) platforms, specifically utilizing training paradigms, instruction tuning, and alignment methods optimized for historical contexts [67]. Technically, the models must be fine-tuned to handle the inconsistent use of long vowels (matres lectionis such as wāw and yā) and the heavy intermixing of Turkic roots with Arabic and Persian vocabulary, ensuring that the system understands the historical spelling variations defining different scribal periods. However, relying heavily on large-scale vision–language models introduces a severe risk of language-prior dominance [67]. Because these deep decoders heavily prioritize statistical linguistic patterns learned during pretraining, they are prone to hallucinations, where linguistic probability overrides the actual visual evidence in the manuscript. To resolve this tension and enforce strict visual grounding, future Chagatai HTR pipelines should move beyond black-box image-to-text generation toward an interpretable framework based on structured prompt composition and cross-attention attribution [68]. By decomposing the transcription task into structured semantic dimensions, such as scribal style, period-specific orthography, and morphological constraints, researchers can utilize structured prompting to guide the multimodal layers more precisely [68]. Furthermore, by calculating token-level spatial relevance scores through mechanisms such as Diffusion Attentive Attribution Maps (DAAM), the system can explicitly map and verify the correspondence between generated text and the original manuscript strokes [68]. This architecture functions as an automated fidelity check by flagging text predictions that lack direct pixel-level support in the image, thereby encouraging the model to capture genuine paleographic variations rather than infer characters solely from modern statistical language priors [67,68].

Stage 5: Open Benchmark and Manuscript-Sensitive Evaluation. The final stage culminates in a reproducible open benchmark that moves beyond isolated experimental setups to standardize evaluation across the field. While standard sequence metrics such as Character Error Rate (CER) and Word Error Rate (WER) remain essential for cross-study comparison, this benchmark introduces manuscript-sensitive protocols for their deployment. Specifically, the evaluation framework establishes error-weighting guidelines where historically acceptable orthographic variations or missing diacritics (nuqta) are tracked separately from catastrophic optical misrecognitions. This prevents minor scribal drift or regional spelling variations from artificially inflating WER penalties. By establishing these standardized, paleographically aware evaluation protocols, the benchmark moves Chagatai HTR toward a cumulative research field that prioritizes the structural and topological invariants of Central Asian intellectual heritage.

Overall, the proposed roadmap suggests that reliable Chagatai HTR cannot emerge from model adaptation alone. It requires a complete research infrastructure in which manuscript digitization, annotation, baseline benchmarking, advanced adaptation, and reproducible evaluation are developed in a coordinated sequence. This staged pathway is summarized in Figure 7.

As shown in Figure 7, the proposed roadmap begins with data foundations and gradually moves toward advanced model adaptation, open benchmarking, and long-term digital access to Chagatai manuscript heritage. Such a staged approach would provide a realistic path toward preserving and computationally accessing Chagatai manuscript heritage while also contributing to broader research on low-resource historical scripts.

5. Limitations of the Study

The current synthesis is subject to several methodological limitations. The review utilized papers identified through selected databases and a predefined search strategy. The search followed PRISMA based steps and covered major academic sources. A significant methodological limitation of this systematic review is the inherent publication bias introduced by the exclusion of non-English regional scholarship. While the search strategy adhered to PRISMA guidelines by targeting major global databases such as Scopus and Web of Science, this approach inevitably overlooks a substantial body of technical and paleographical research published in Kazakh, Uzbek, and Russian. In the specific context of Central Asian digital heritage, critical primary research and specialized algorithmic adaptations are frequently disseminated through regional academic journals and institutional archival reports that remain unindexed in Western-centric platforms. Consequently, the validity of the Proposed Roadmap (Section 4) may be constrained by the omission of localized paleographical expertise, including specific codicological insights into Samarkand rag paper aging, regional scribal variations, and alternative computational frameworks developed within Central Asian technical universities. This linguistic and digital barrier represents a recognized systemic challenge in digital humanities; by relying on English-indexed literature, the synthesis potentially prioritizes global architectural trends over indigenous engineering solutions tailored to the unique orthographic nuances of the Eastern Turki tradition. Acknowledging this bias is essential for maintaining research validity, as it highlights that the current state of the art, while robust in the global sphere, remains partially decoupled from the deep-seated regional scholarship necessary for a truly comprehensive Chagatai HTR infrastructure.

The review examined studies published between 1 January 2020 and 1 March 2026. This period reflects recent progress in deep learning, Transformer architectures, multimodal learning, and document analysis. The selected publication window places particular attention on recent AI-driven recognition methods. As a result, earlier contributions in classical OCR, paleography, and manuscript digitization receive less attention within the review.

Building on these constraints, the paucity of primary Chagatai-specific HTR literature necessitated a reliance on methodological proxies from related Perso–Arabic and Turkic-script domains. While analysis of Ottoman Turkish, Persian, and Kazakh traditions offers valuable methodological patterns, their findings remain contingent upon further script-specific validation to account for Chagatai’s unique orthographic nuances.

A secondary limitation concerns the heterogeneous degree of peer-review rigor across the included studies. Approximately 16% of the reviewed corpus consists of preprints (N = 8), which, while technically robust, have not undergone formal academic validation. This introduces a variable level of confidence in certain reported performance metrics. To mitigate this risk, the synthesis prioritizes architectural trends and methodological innovations from preprints rather than treating their empirical results as absolute field-level benchmarks, ensuring that the study’s conclusions remain grounded in a balanced evidence base.

Methodological variation across the reviewed studies limits direct comparison of reported results. The studies included in the review rely on diverse evaluation metrics, such as Character Error Rate (CER) and mean Average Precision (mAP). This methodological variation restricts direct comparison of results and hinders the objective assessment of competing architectures. Data availability and dataset quality present another challenge. Many studies rely on private collections, synthetic datasets, or small archival samples. Public benchmark datasets remain scarce. This situation restricts reproducibility and limits independent validation of published recognition performance.

This study is implemented as a systematic literature review and indirect evidence synthesis. The scope focuses on the synthesis of existing research, the identification of research gaps, and the definition of strategic research directions. The primary contribution lies in mapping the research landscape and establishing a foundational development pathway for Chagatai handwritten text recognition, serving as a conceptual framework to guide subsequent empirical validation.

6. Conclusions

This study presented a systematic review of artificial intelligence models and methods for handwritten text recognition and restoration in low-resource and historical scripts, with a specific focus on the future development of Chagatai manuscript recognition. The review synthesized recent studies published between 2020 and 2026 and grouped the literature according to recognition methods, script and character classification, system-level recognition, learning and adaptation strategies, restoration and enhancement, data creation, benchmarking, multimodal learning, GAN-based augmentation, and architectural trends.

Empirical evidence indicates a transformative shift from modular, segmentation-dependent pipelines toward unified, attention-driven architectures, including Transformer-based, Vision Transformer-based, and multimodal vision–language systems. However, CNN-based and CNN–RNN models remain important as stable and practical baselines, particularly in low-resource settings. The review also shows that multimodal systems can improve recognition and restoration performance, but their success depends on reliable alignment between visual evidence and linguistic context.

For Chagatai, the review shows the main barrier is not just the lack of a specialized recognition model. The deeper problem is the absence of full research infrastructure. This includes a representative digitized corpus, expert checked transcriptions, standard annotation rules, preprocessing protocols, and open benchmarks. Without these resources, advanced models cannot be evaluated or compared in a reliable way.

The study identifies several foundational directions for the field. The development of reliable Chagatai HTR necessitates a coordinated, multi-stage research trajectory that integrates paleography with computational infrastructure. This trajectory commences with the establishment of corpus building and benchmark design as the immediate empirical priorities. Following these foundational steps, the research program will transition to the systematic testing of baseline HTR models, advanced Transformer architectures, multimodal systems, and restoration methods. This staged pathway provides a realistic framework for the digital preservation and computational reintegration of Central Asian intellectual heritage into the global scholarly record.