Bangla-SGP

Bangla-SGP is a parallel dataset of human-annotated and rule-augmented Bangla sentence–gloss pairs for Bangla Sign Language (BdSL) translation research: 997 human-annotated pairs (annotated by a professional signer) plus 3,000 synthetically generated pairs (rule-based morphological transformation, masked-token substitution, and RAG-based few-shot prompting released with a GPT-4.1-nano two-stage pipeline).

Paper: Introducing A Bangla Sentence – Gloss Pair Dataset for Bangla Sign Language Translation and Research (LREC 2026, pp. 10457–10466, landing page, DOI).

Code and full documentation: github.com/OpenCR-lab/bangla-sgp.

Why: the creation rationale

Gloss sequences — individual Bangla sign–supported words in signed order — serve as an intermediate representation for continuous signing. Bangladesh has ~13 million people with hearing difficulties (3 million deaf); research on sentence-level BdSL was blocked by the lack of a sentence-level dataset. Bangla-SGP supplies that layer, annotated by a professional signer.

Configuration and structure

Two configurations share the same validation and test files (the original 1,000-pair set was split 80/10/10; identical val/test are used both times so results with and without augmentation stay comparable):

config	train	validation	test	what it is
human (default)	data/human/train.json	data/val.json	data/test.json	expert-annotated pairs only (997)
augmented	data/augmented/train.json	data/val.json	data/test.json	the shared human train rows + 3,000 synthetic pairs (3,797)

Each row: id (stable ID), sentence (Bangla), gloss (BdSL gloss sequence, Bangla script), topic, subtopic, tense, type, generation_type (human_annotation / morph_aug / context_aug / rag_aug). Example:

{"id": "bsgp-h-0001",
 "sentence": "প্লাবিত গ্রামে মোবাইল মেডিকেল টিম পাঠানো হয়েছে।",
 "gloss": "পানির প্লাবিত গ্রাম ওখানে মোবাইল মেডিকেল টিম যাওয়া হয়েছে।",
 "topic": "health_and_disease", "subtopic": "mobile_medical_team",
 "tense": "present_perfect", "type": "simple",
 "generation_type": "human_annotation"}

Augmentation methods (summary)

1. Rule-based Morphological Transformation (morph_aug, 498): tense rules compiled with the professional signer (e.g., future: verb → nominalised verb + "হবে").
2. Masked-Token Substitution (context_aug, 383): BanglaBERT fills masked nouns/verbs inside rule-safe templates to build lexical variants.
3. RAG-based few-shot prompting (rag_aug, 2,119): LaBSE embeddings + Pinecone retrieval (threshold 0.5, fallback to rules-only), two-stage prompting with GPT-4.1-nano; published validation 75.3% validation rate, quality 3.19/5, binary κ = 0.7489.

Full rules and prompt templates: docs/prompts/translation-rules-prompt.md and docs/linguistic-rules.md. Datasheet: DATASHEET.md.

Results on this dataset (from the paper, Tables 2–3)

Model	Dataset	BLEU-1	BLEU-2	BLEU-3	BLEU-4	COMET
mT5-small	Base (1K)	51.79	36.26	25.73	17.56	0.8564
mT5-small	Augmented (4K)	55.33	40.11	29.06	20.28	0.8602
mBART-50	Base (1K)	56.65	40.96	29.36	21.06	0.7829
mBART-50	Augmented (4K)	61.09	46.01	35.09	27.31	0.8261
GPT-4.1-nano	Augmented (4K)	57.11	41.50	31.11	22.42	0.8913
NLLB-200-1.3B	Base (1K)	67.99	40.82	26.49	16.67	0.907

Reference-only side-by-side with RWTH-PHOENIX-2014T is printed in the paper with the explicit note that direct cross-linguistic comparison of BLEU scores is not implied.

Limitations & considerations

• Text-only. No video, audio, pose, or non-manual markers; glosses cannot represent facial expression or body movement. Phase 2 (synchronized video + non-manual markers) is planned (paper §7.2).
• Single base annotator — the paper flags potential signer bias.
• No dedicated handling of names/landmarks (no FS-NAME-style tokens, no fingerspelled sequences).
• Signer credit: Mr. Ariful Islam (BTV) annotated; Mr. Ariful Islam and Mrs. Tanjila Tartushi (BTV) validated independently — with the annotators' informed consent (paper §7.3).
• Counts: paper round-phrases are "1,000 / ~3,000 / 4,000"; the released files carry 997 / 3,000 / 3,997 rows. Seven human sentence strings recur; three synthetic pairs repeat (details, §6).

Licence

Data: CC BY-4.0. Repository code and documentation: MIT. Attribution requirements: cite the paper and the dataset; see DATA_LICENSE.md.

Citation

See the repository's CITATION.cff for the machine-readable form; BibliTeX:

@inproceedings{saha-etal-2026-introducing,
  title     = "Introducing A Bangla Sentence -- Gloss Pair Dataset for
               Bangla Sign Language Translation and Research",
  author    = {Saha, Neelavro and Shahriyar, Rafi and Roudra, Nafis Ashraf and
               Sakib, Saadman and Rasel, Annajiat Alim},
  booktitle = {Proceedings of the Fifteenth Language Resources and Evaluation
               Conference (LREC 2026)},
  pages     = {10457--10466},
  year      = {2026},
  publisher = {European Language Resources Association (ELRA)},
  doi       = {10.63317/38qenrwzegr9},
  url       = {https://lrec.elra.info/lrec2026-main-820}
}

Questions: the organisation's contact on the GitHub repository README.