NeelavroSaha commited on
Commit
4b1ffd2
·
verified ·
1 Parent(s): 5a905b8

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +161 -0
README.md ADDED
@@ -0,0 +1,161 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ pretty_name: Bangla-SGP
3
+ license: cc-by-4.0
4
+ language:
5
+ - bn
6
+ task_categories:
7
+ - translation
8
+ tags:
9
+ - sign-language
10
+ - gloss
11
+ - bangla
12
+ - bdsl
13
+ - low-resource
14
+ - data-augmentation
15
+ size_categories:
16
+ - 1K<n<10K
17
+ source_datasets: []
18
+ configs:
19
+ - config_name: human
20
+ default: true
21
+ data_files:
22
+ - split: train
23
+ path: data/human/train.json
24
+ - split: validation
25
+ path: data/val.json
26
+ - split: test
27
+ path: data/test.json
28
+ - config_name: augmented
29
+ data_files:
30
+ - split: train
31
+ path: data/augmented/train.json
32
+ - split: validation
33
+ path: data/val.json
34
+ - split: test
35
+ path: data/test.json
36
+ ---
37
+
38
+ # Bangla-SGP
39
+
40
+ **Bangla-SGP** is a parallel dataset of human-annotated and rule-augmented
41
+ Bangla sentence–gloss pairs for **Bangla Sign Language (BdSL) translation
42
+ research**: 997 human-annotated pairs (annotated by a professional signer)
43
+ plus 3,000 synthetically generated pairs (rule-based morphological
44
+ transformation, masked-token substitution, and RAG-based few-shot
45
+ prompting released with a GPT-4.1-nano two-stage pipeline).
46
+
47
+ Paper: *Introducing A Bangla Sentence – Gloss Pair Dataset for Bangla Sign
48
+ Language Translation and Research* (LREC 2026, pp. 10457–10466,
49
+ [landing page](https://lrec.elra.info/lrec2026-main-820),
50
+ [DOI](https://doi.org/10.63317/38qenrwzegr9)).
51
+
52
+ Code and full documentation: [github.com/OpenCR-lab/bangla-sgp](https://github.com/OpenCR-lab/bangla-sgp).
53
+
54
+ ## Why: the creation rationale
55
+
56
+ Gloss sequences — individual Bangla sign–supported words in signed order —
57
+ serve as an intermediate representation for continuous signing. Bangladesh
58
+ has ~13 million people with hearing difficulties (3 million deaf); research
59
+ on sentence-level BdSL was blocked by the lack of a sentence-level dataset.
60
+ Bangla-SGP supplies that layer, annotated by a professional signer.
61
+
62
+ ## Configuration and structure
63
+
64
+ Two configurations share the same validation and test files (the original
65
+ 1,000-pair set was split 80/10/10; identical val/test are used both times so
66
+ results with and without augmentation stay comparable):
67
+
68
+ | config | train | validation | test | what it is |
69
+ |---|---|---:|---:|---|
70
+ | `human` (default) | `data/human/train.json` | `data/val.json` | `data/test.json` | expert-annotated pairs only (997) |
71
+ | `augmented` | `data/augmented/train.json` | `data/val.json` | `data/test.json` | the shared human train rows + 3,000 synthetic pairs (3,797) |
72
+
73
+ Each row: `id` (stable ID), `sentence` (Bangla), `gloss` (BdSL gloss
74
+ sequence, Bangla script), `topic`, `subtopic`, `tense`, `type`,
75
+ `generation_type` (`human_annotation` / `morph_aug` / `context_aug` /
76
+ `rag_aug`). Example:
77
+
78
+ ```json
79
+ {"id": "bsgp-h-0001",
80
+ "sentence": "প্লাবিত গ্রামে মোবাইল মেডিকেল টিম পাঠানো হয়েছে।",
81
+ "gloss": "পানির প্লাবিত গ্রাম ওখানে মোবাইল মেডিকেল টিম যাওয়া হয়েছে।",
82
+ "topic": "health_and_disease", "subtopic": "mobile_medical_team",
83
+ "tense": "present_perfect", "type": "simple",
84
+ "generation_type": "human_annotation"}
85
+ ```
86
+
87
+ ## Augmentation methods (summary)
88
+
89
+ 1. **Rule-based Morphological Transformation** (`morph_aug`, 498): tense
90
+ rules compiled with the professional signer (e.g., future: verb →
91
+ nominalised verb + "হবে").
92
+ 2. **Masked-Token Substitution** (`context_aug`, 383): BanglaBERT fills
93
+ masked nouns/verbs inside rule-safe templates to build lexical variants.
94
+ 3. **RAG-based few-shot prompting** (`rag_aug`, 2,119): LaBSE embeddings +
95
+ Pinecone retrieval (threshold 0.5, fallback to rules-only), two-stage
96
+ prompting with GPT-4.1-nano; published validation 75.3% validation rate,
97
+ quality 3.19/5, binary κ = 0.7489.
98
+
99
+ Full rules and prompt templates: [`docs/prompts/translation-rules-prompt.md`](https://github.com/OpenCR-lab/bangla-sgp/blob/main/docs/prompts/translation-rules-prompt.md)
100
+ and [`docs/linguistic-rules.md`](https://github.com/OpenCR-lab/bangla-sgp/blob/main/docs/linguistic-rules.md).
101
+ Datasheet: [`DATASHEET.md`](https://github.com/OpenCR-lab/bangla-sgp/blob/main/DATASHEET.md).
102
+
103
+ ## Results on this dataset (from the paper, Tables 2–3)
104
+
105
+ | Model | Dataset | BLEU-1 | BLEU-2 | BLEU-3 | BLEU-4 | COMET |
106
+ |---|---|---:|---:|---:|---:|---:|
107
+ | mT5-small | Base (1K) | 51.79 | 36.26 | 25.73 | 17.56 | 0.8564 |
108
+ | mT5-small | Augmented (4K) | 55.33 | 40.11 | 29.06 | 20.28 | 0.8602 |
109
+ | mBART-50 | Base (1K) | 56.65 | 40.96 | 29.36 | 21.06 | 0.7829 |
110
+ | mBART-50 | Augmented (4K) | 61.09 | 46.01 | 35.09 | 27.31 | 0.8261 |
111
+ | GPT-4.1-nano | Augmented (4K) | 57.11 | 41.50 | 31.11 | 22.42 | 0.8913 |
112
+ | NLLB-200-1.3B | Base (1K) | 67.99 | 40.82 | 26.49 | 16.67 | 0.907 |
113
+
114
+ Reference-only side-by-side with RWTH-PHOENIX-2014T is printed in the paper
115
+ with the explicit note that direct cross-linguistic comparison of BLEU
116
+ scores is not implied.
117
+
118
+ ## Limitations & considerations
119
+
120
+ - **Text-only.** No video, audio, pose, or non-manual markers; glosses cannot
121
+ represent facial expression or body movement. Phase 2 (synchronized video +
122
+ non-manual markers) is planned (paper §7.2).
123
+ - **Single base annotator** — the paper flags potential signer bias.
124
+ - **No dedicated handling of names/landmarks** (no FS-NAME-style tokens, no
125
+ fingerspelled sequences).
126
+ - **Signer credit:** Mr. Ariful Islam (BTV) annotated; Mr. Ariful Islam and
127
+ Mrs. Tanjila Tartushi (BTV) validated independently — with the annotators'
128
+ informed consent (paper §7.3).
129
+ - Counts: paper round-phrases are "1,000 / ~3,000 / 4,000"; the released
130
+ files carry 997 / 3,000 / 3,997 rows. Seven human sentence strings recur;
131
+ three synthetic pairs repeat
132
+ ([details](https://github.com/OpenCR-lab/bangla-sgp/blob/main/docs/dataset-statistics.md),
133
+ §6).
134
+
135
+ ## Licence
136
+
137
+ Data: **CC BY-4.0**. Repository code and documentation: **MIT**. Attribution
138
+ requirements: cite the paper and the dataset; see `DATA_LICENSE.md`.
139
+
140
+ ## Citation
141
+
142
+ See the repository's [CITATION.cff](https://github.com/OpenCR-lab/bangla-sgp) for the
143
+ machine-readable form; BibliTeX:
144
+
145
+ ```bibtex
146
+ @inproceedings{saha-etal-2026-introducing,
147
+ title = "Introducing A Bangla Sentence -- Gloss Pair Dataset for
148
+ Bangla Sign Language Translation and Research",
149
+ author = {Saha, Neelavro and Shahriyar, Rafi and Roudra, Nafis Ashraf and
150
+ Sakib, Saadman and Rasel, Annajiat Alim},
151
+ booktitle = {Proceedings of the Fifteenth Language Resources and Evaluation
152
+ Conference (LREC 2026)},
153
+ pages = {10457--10466},
154
+ year = {2026},
155
+ publisher = {European Language Resources Association (ELRA)},
156
+ doi = {10.63317/38qenrwzegr9},
157
+ url = {https://lrec.elra.info/lrec2026-main-820}
158
+ }
159
+ ```
160
+
161
+ Questions: the organisation's contact on the GitHub repository README.