The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
SSLGen Sinhala–Gloss Sentence Dataset
This dataset contains a teacher-verified Sinhala sentence–gloss parallel corpus developed as part of the SSLGen: Sinhala Sign Language Text-to-Gloss-to-Video Translation Framework.
The corpus was created to support research on Sinhala Sign Language (SSL) translation, particularly the task of converting written Sinhala sentences into Sinhala Sign Language gloss sequences.
The dataset contains 3,363 Sinhala sentence–gloss pairs collected from selected Grade 1–9 Sinhala school textbooks and verified by a native Sinhala Sign Language teacher.
Dataset Overview
| Property | Description |
|---|---|
| Dataset Name | SSLGen Sinhala–Gloss Sentence Dataset |
| Language | Sinhala |
| Target Representation | Sinhala Sign Language Gloss |
| Number of Sentence–Gloss Pairs | 3,363 |
| Source | Selected Grade 1–9 Sinhala textbooks |
| Verification | Native Sinhala Sign Language teacher |
| Primary Task | Sinhala Text-to-Gloss Translation |
| Main Model | NLLB-200-distilled-600M |
| Baseline | Sinhala BERT Medium-V2 |
| Evaluation Metric | BLEU-4 |
Dataset Description
The dataset consists of paired Sinhala sentences and their corresponding Sinhala Sign Language gloss representations.
Each example contains:
Sinhala Sentence → SSL Gloss Sequence
For example:
Sinhala sentence
↓
Corresponding SSL gloss sequence
The gloss representation provides an intermediate linguistic representation between written Sinhala and the sign video resources used by the SSLGen framework.
Data Collection
The sentences were collected from Grade 1–9 Sinhala school textbooks.
The textbook material was reviewed to identify sentences suitable for constructing a Sinhala–SSL gloss corpus.
Certain textbook content was excluded when it was considered unsuitable for the intended sentence-level translation task, including:
- Poems
- Older/archaic Sinhala stories and language
- Grammar-focused sections
- Other content that did not provide suitable sentence-level material for the target translation task
The resulting sentence–gloss pairs were reviewed and verified by a native Sinhala Sign Language teacher to improve the linguistic validity of the gloss annotations.
Dataset Statistics
The final corpus contains:
3,363 Sinhala sentence–gloss pairs
The corpus was developed specifically to address the limited availability of publicly accessible Sinhala–Sinhala Sign Language parallel resources.
Because Sinhala Sign Language is a relatively low-resource language, the corpus is intended to provide a foundation for future research in:
- Sinhala Sign Language translation
- Low-resource machine translation
- Sign language NLP
- Sinhala NLP
- Text-to-gloss translation
- Gloss-based sign language generation
Dataset Structure
The dataset contains paired sentence and gloss information.
A typical example follows the structure:
{
"sentence": "අම්මා ආහාර උයනවා",
"gloss": "අම්මා කෑම උයනවා"
}
Main Fields
| Field | Description |
|---|---|
sentence |
Written Sinhala input sentence |
gloss |
Corresponding Sinhala Sign Language gloss sequence |
Gloss Representation
Glosses are used as an intermediate representation between Sinhala text and sign video generation.
The overall SSLGen pipeline is:
Sinhala Text
│
▼
Sinhala → SSL Gloss Translation
│
│
▼
Predicted Gloss Sequence
│
▼
Gloss-to-Video Retrieval
│
▼
Sinhala Sign Language Video
The gloss representation allows the text-to-gloss translation task to be evaluated separately from the subsequent video generation stage.
Research Methodology
The corpus was used to fine-tune a multilingual neural machine translation model for Sinhala-to-gloss translation.
Primary Model
NLLB-200-distilled-600M
The model was fine-tuned on the Sinhala–gloss parallel corpus using language-specific training data.
Baseline
Sinhala BERT Medium-V2
A Sinhala-specific pretrained language model was used as a baseline for comparison.
Evaluation
The models were evaluated using BLEU-4.
The reported test-set results were:
| Model | Test BLEU-4 |
|---|---|
| NLLB-200-distilled-600M | 46.14 |
| Sinhala BERT Medium-V2 | 35.46 |
The fine-tuned NLLB-200 model therefore achieved higher BLEU-4 performance than the Sinhala BERT Medium-V2 baseline on the evaluated test set.
Intended Use
This dataset is intended primarily for academic and research purposes.
Potential applications include:
- Sinhala-to-gloss machine translation
- Sinhala Sign Language translation
- Low-resource neural machine translation
- Sign language NLP
- Sinhala NLP
- Gloss prediction
- Text-to-sign language research
- Accessibility technology
- Multilingual translation research
- Sign language generation pipelines
Researchers may use the corpus to investigate different neural architectures, pretrained multilingual models, sequence-to-sequence models, and other approaches for low-resource translation.
Limitations
The dataset has several important limitations.
Limited Size
The corpus contains 3,363 sentence–gloss pairs, which is relatively small compared with large-scale machine translation datasets.
Domain Limitation
The sentences were collected from Grade 1–9 Sinhala textbooks. Therefore, the corpus may not fully represent everyday Sinhala usage, conversational language, news language, or other real-world domains.
Linguistic Coverage
The dataset does not represent the complete vocabulary or grammatical structure of Sinhala Sign Language.
Gloss-Based Representation
Glosses are an intermediate representation and cannot fully capture all linguistic properties of sign language, including:
- Facial expressions
- Body posture
- Spatial referencing
- Non-manual markers
- Timing and movement characteristics
- Other signer-specific linguistic features
Therefore, the gloss representation should not be considered a complete representation of Sinhala Sign Language.
Relationship to the SSLGen Video Dataset
This sentence dataset forms the text-to-gloss component of the SSLGen framework.
It is used together with the SSLGen video resources for the complete translation pipeline:
SSLGen Framework
┌──────────────────────────┐
│ Sinhala Sentence │
└────────────┬─────────────┘
│
▼
┌──────────────────────────┐
│ Sinhala → SSL Gloss │
│ NLLB-200 │
└────────────┬─────────────┘
│
▼
┌──────────────────────────┐
│ Predicted Gloss │
└────────────┬─────────────┘
│
▼
┌──────────────────────────┐
│ Gloss-to-Video │
│ Retrieval │
└────────────┬─────────────┘
│
▼
┌──────────────────────────┐
│ Sinhala Sign Language │
│ Video │
└──────────────────────────┘
The retrieval-based video generation component does not require sentence-level paired text-to-video training data.
Data Quality and Verification
The Sinhala–gloss pairs were reviewed and verified by a native Sinhala Sign Language teacher.
Teacher verification was used to improve the correspondence between the written Sinhala sentences and their intended Sinhala Sign Language gloss representations.
However, teacher verification does not imply that the dataset represents every regional, contextual, or signer-specific variation of Sinhala Sign Language.
Source and Attribution
The sentence material was derived from selected Grade 1–9 Sinhala school textbooks.
This dataset represents a research corpus constructed from selected source material and annotated with Sinhala Sign Language glosses.
Users should appropriately acknowledge the underlying educational source materials when required and should ensure that their use of the dataset complies with any applicable rights or restrictions associated with the source materials.
The gloss annotations and corpus construction were developed as part of the SSLGen research project.
License and Usage
This dataset is provided for research and academic use.
Before redistributing, commercially using, or creating derivative datasets from the sentence material, users should verify that such use is permitted for the underlying source material from which the sentences were derived.
The copyright status of the original textbook material may be separate from the rights associated with the newly created gloss annotations and dataset organization.
Users are responsible for ensuring that their use of individual sentence content complies with applicable copyright and other legal requirements.
Citation
If you use the SSLGen Sinhala–Gloss Sentence Dataset in your research, please cite the SSLGen research project:
@misc{sslgen2026,
title = {SSLGen: Sinhala Sign Language Text-to-Gloss-to-Video Translation Framework},
author = {Geethma Yasashwi Wickramarathne and Nemasha Hirushi},
year = {2026},
note = {Sinhala Sign Language text-to-gloss-to-video translation research project}
}
If you use sentences derived from specific textbooks, please also acknowledge and cite the relevant source materials where appropriate.
Acknowledgements
We would like to thank the native Sinhala Sign Language teacher who contributed to the verification of the sentence–gloss pairs.
We also acknowledge the educational resources from which the source sentences were selected.
We thank everyone who supported the collection, annotation, verification, and development of this corpus.
Disclaimer
This dataset is intended to support research into Sinhala Sign Language and low-resource language technologies.
The dataset does not claim to represent the complete Sinhala Sign Language lexicon, grammar, or linguistic variation.
The gloss representation is an intermediate computational representation and should not be interpreted as a complete linguistic description of Sinhala Sign Language.
Users are responsible for ensuring that their use of the dataset complies with applicable copyright, licensing, and other legal requirements concerning the underlying source material.
Related Resources
Fine-Tuned NLLB-200 Model
The fine-tuned NLLB-200 model developed using this corpus is available at:
https://huggingface.co/GeethmaYasashwi/NLLB-Fine-tuned
SSLGen Video Dataset
The corresponding Sinhala Sign Language video resources are available at:
https://huggingface.co/datasets/GeethmaYasashwi/SSL-Gen_Video-Dataset
Project
SSLGen: Sinhala Sign Language Text-to-Gloss-to-Video Translation Framework
Researchers: Geethma Yasashwi Wickramarathne Nemasha Hirushi
Year: 2026
- Downloads last month
- 18