Sign Language Recognition and Translation for Low-Resource Languages: Challenges and Pathways Forward

Nigar Alishzade Gulchin Abdullayeva · 2026

Read Sign Language Recognition and Translation for Low-Resource Languages: Challenges and Pathways Forward by Nigar Alishzade Gulchin Abdullayeva (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-12096.txt.

Introduction

Nigar Alishzade and Gulchin Abdullayeva review, in 2026 work, sign language reading and signing back for low-resource tongues, using AzSL as a case study. Over 300 sign tongues remain low-resource from thin records, sparse data, and few digital tools. Their review draws eight lessons, including local joint design, dialect variety capture, and privacy-safe pose pictures. They propose shifts from model-centric to data-centric AI, signer-free to signer-tuned systems, and gold-standard to task-based metrics, with a MediaPipe-based roadmap for AzSL.

Catalog record

CREATORNigar Alishzade Gulchin Abdullayeva
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE47617
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.12096

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Sign Language Recognition and Translation for Low-Resource Languages: Challenges and Pathways Forward Nigar Alishzade Gulchin Abdullayeva Abstract Sign languages are natural, visual-gestural languages used by Deaf communities worldwide. Over 300 distinct sign languages remain severely low-resource due to limited documentation, sparse datasets, and insufficient computational tools. This systematic review synthesizes literature on sign language recognition and translation for under-resourced languages, using Azerbaijan Sign Language (AzSL) as a case study. We identify seven interconnected challenges—data scarcity, annotation bottlenecks, non-manual feature neglect, inter-class similarity, recognition errors, translation failures, and poor generalization—that create a self-reinforcing cycle impeding low-resource development. Analysis of global initiatives extracts eight actionable lessons, including community co-design, dialectal diversity capture, and privacy-preserving pose-based representations. Turkic sign languages (Kazakh, Turkish, Azerbaijani) receive special attention, as linguistic proximity enables effective transfer learning. We propose three paradigm shifts: from architecture-centric to data-centric AI, from signer-independent to signer-adaptive systems, and from reference-based to task-specific evaluation metrics. Recommendations align with UN Sustainable Development Goals 4, 5, 10, and 17. A technical roadmap for AzSL leverages lightweight MediaPipe-based architectures, community-validated annotations, and offline-first deployment. Progress requires sustained interdisciplinary collaboration centered on Deaf communities to ensure cultural authenticity, ethical governance, and practical communication benefit. 1 Introduction Sign languages are natural, visual-gestural languages used by Deaf communities worldwide. Over 300 distinct sign languages exist globally, yet the vast majority remain low-resource due to limited documentation, sparse datasets, and a lack of dedicated computational tools [ 1 , 2 ] . This digital divide stands in stark contrast to rapid progress in spoken language processing. In Azerbaijan, AzSL serves approximately 30,000 Deaf individuals but only recently gained official recognition (2025) and maintains a nascent digital presence, exemplified by the AzSLD dataset of isolated signs [ 3 ] . SLR identifies signs from video input, while SLT extends this to generate spoken or written language equivalents. Recent advances in deep learning—employing 3D CNNs, Vision Transformers, and graph neural networks—have achieved isolated SLR accuracy exceeding 98% on high-resource benchmarks such as WLASL and PHOENIX-2014T [ 4 ] . However, these results depend critically on large, well-annotated corpora. When applied to low-resource sign languages, state-of-the-art models exhibit substantial performance degradation, with error rates commonly 10–30% higher than for high-resource counterparts [ 1 ] . Contributions and Paper Structure This review makes three primary contributions. First, we provide systematic synthesis of recent literature (2019–2026) with emphasis on how challenges interact to disproportionately hinder low-resource sign language development. Second, we conduct comparative analysis of international initiatives and, using AzSL as a representative case, develop a contextually-grounded roadmap that balances technical innovation with community engagement and ethical considerations. Third, we explicitly align our recommendations with United Nations Sustainable Development Goals to situate sign language technology within broader development and equity frameworks. The paper proceeds as follows: Section 2 describes the systematic review methodology. Section 3 analyzes the standard SLR/SLT pipeline to identify failure points in low-resource contexts. Section 4 synthesizes critical challenges and their interdependencies. Section 5 presents comparative analysis of global initiatives with strengthened focus on Central Asian and Turkic language initiatives. Section 6 discusses strategic implications, paradigm shifts, and SDG alignment. Section 7 provides the technical analysis and roadmap, and Section 8 concludes with recommendations. 2 Methodology This review was conducted following PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines for scoping reviews to ensure transparency and reproducibility. Search Strategy We searched five electronic databases for peer-reviewed articles, conference proceedings, and preprints published between January 2019 and February 2026: IEEE Xplore, ACM Digital Library, arXiv, Scopus, and the ACL Anthology. The search strategy combined three concept groups: 1. Sign language systems: “sign language,” “SLR,” “SLT,” “gesture recognition” 2. Resource constraints: “low-resource,” “under-resourced,” “data scarcity,” “small dataset” 3. Geographic/demographic context: “developing country,” “minority language,” “endangered language,” “Turkic,” “Central Asia” These terms were combined using Boolean operators to identify relevant studies. Inclusion and Exclusion Criteria Inclusion criteria: (1) primary focus on automatic sign language recognition or translation; (2) explicit discussion of challenges related to data availability, annotation, or model generalization; (3) publication in English; (4) peer-reviewed venue or substantial preprint with clear methodology. Exclusion criteria: (1) focus solely on gesture recognition for human-computer interaction without sign language linguistic context; (2) duplicate publications; (3) insufficient technical detail; (4) review papers without primary research contribution (though some reviews were consulted for reference purposes). Study Selection and Data Extraction The initial search yielded 412 records. Following PRISMA procedures (illustrated in Figure 1 ), we removed 125 duplicates and screened 287 unique records based on title and abstract. Full-text assessment of 103 articles resulted in inclusion of 34 studies in the qualitative synthesis. Data were extracted into a structured table recording: study characteristics, sign language(s) examined, dataset size, technical approaches, reported challenges, and relevance to low-resource contexts. Due to heterogeneity in evaluation protocols and inconsistent metrics across studies, quantitative meta-analysis was not feasible; findings are synthesized narratively with illustrative quantitative comparisons where applicable. Identification Records identified from 5 databases (IEEE Xplore, ACM, arXiv, Scopus, ACL) n = 412 Records after duplicate removal n = 287 Screening Records screened (title & abstract) n = 287 Records excluded n = 184 Eligibility Full-text articles assessed n = 103 Full-text articles excluded n = 69 Included Studies in qualitative synthesis n = 34 Figure 1: PRISMA 2020 flow diagram of the study selection process. Of 412 identified records, 34 studies met inclusion criteria and were included in qualitative synthesis. 3 Pipeline Analysis Through a Low-Resource Lens Before synthesizing aggregated challenges, we examine standard SLR/SLT pipeline components to understand where and why low-resource conditions cause systematic failures. Input Acquisition and Preprocessing High-resource pipelines typically assume multi-camera studio setups with controlled lighting and static backgrounds. In contrast, low-resource pipelines must contend with uncontrolled conditions: variable smartphone video capture, occlusions, low resolution, and diverse backgrounds [ 3 ] . Preprocessing steps designed to isolate hands and face often fail when signers wear traditional attire, jewelry, or when recording occurs outside studio environments. This preprocessing fragility directly amplifies downstream errors. Feature Extraction and Transfer Learning Current practice commonly employs transfer learning from models pre-trained on large action recognition datasets (e.g., Kinetics) or high-resource sign languages (e.g., ASL). While effective for capturing manual features, this approach assumes feature transferability across linguistic and cultural contexts. Linguistic analysis challenges this assumption: sign languages differ significantly in phonological structure and in the grammatical function of non-manual markers [ 5 ] . A 3D CNN pre-trained on ASL may prioritize features that carry no linguistic meaning in AzSL or that conflict with AzSL linguistic structure. Temporal Modeling and Continuous Recognition For continuous SLR, Connectionist Temporal Classification (CTC) enables unsegmented sequence-to-gloss alignment without explicit word boundaries. However, low-resource models struggle with co-articulation effects (smoothing transitions between signs) and signer variability, as limited training data cannot capture the full distribution of sign execution speeds and transitional patterns. Consequently, word error rates (WER) remain prohibitively high. Gloss-free approaches that generate direct text output are similarly constrained due to their substantially higher data requirements [ 4 ] . Output Generation and Evaluation Metrics Post-processing and translation models require parallel corpora of sign glosses (linguistic notation) and spoken language text. Such corpora are virtually non-existent for low-resource languages. Moreover, standard metrics—WER and BLEU—inadequately measure

Read the complete text