zahemen9900/signtalk-gh-sampled

SignTalk-GH + GSL Dictionary (Sampled) — Dataset Card for Hugging Face

Dataset generated by merging the SignTalk-GH corpus with the GSL dictionary (zenodo openpose videos) and sampling to a balanced subset. Intended for downstream preprocessing (preprocessing.ipynb) and Text2Sign pipelines.

Summary

• Source corpora: SignTalk-GH (Videos + Metadata.xlsx) and GSL OpenPose data.
• Integration: GSL concepts normalized (token splitting on _OR_, _AND_, hyphens preserved), de-duped against existing SignTalk sentences, and copied into the SignTalk video layout.
• Sampling: Stratified by category with per-sentence variant caps; target size ~1,500 videos (or ~2,500 if DOUBLE_SAMPLE=True).
• Artifacts:
  • sample_dataset/videos/*.mp4
  • sample_dataset/sampled_metadata.csv
  • sample_dataset/dataset_summary.csv
  • Optional zip produced in Colab: SignTalk-GH_Sampled.zip

Data Fields (sampled_metadata.csv)

• video_file (str): File name of the mp4 in videos/ (e.g., 123A.mp4).
• sentence_id (int): Unique sentence identifier from SignTalk; GSL additions start after the max existing ID.
• sentence (str): Normalized sentence/gloss text.
• category (str): Original SignTalk category or GSL Dictionary for imported entries.
• variant (str): Variant token inferred from filename suffix; BASE if none.

Directory Structure

sample_dataset/
├── videos/                # mp4 clips
├── sampled_metadata.csv   # tabular metadata
└── dataset_summary.csv    # quick stats

Creation Steps (notebook key points)

1. Download sources: SignTalk-GH zip (Kaggle) + GSL OpenPose zip (Zenodo).
2. Parse GSL folders, normalize concept tokens, and copy variants into SignTalk Videos/ with new IDs.
3. Map videos to sentences via filename patterns (e.g., 1A.mp4).
4. Stratified sampling by category with per-sentence variant bounds (MIN_VARIANTS_PER_SENTENCE, MAX_VARIANTS_PER_SENTENCE).
5. Copy sampled videos into sample_dataset/videos/ and write sampled_metadata.csv + dataset_summary.csv.

Example Row

video_file: "184C.mp4"
sentence_id: 184
sentence: "Go To Hospital"
category: "Healthcare"
variant: "C"

Loading with datasets.load_dataset (after pushing to HF Hub)

from datasets import load_dataset

ds = load_dataset("zahemen9900/signtalk-gh-gsl", split="train")
print(ds)
print(ds[0])  # columns: video_file, sentence_id, sentence, category, variant

If storing videos as files in the repo, set keep_in_memory=False and ensure videos/ paths are relative. For large media, prefer an external storage bucket and include URLs in video_file.

Recommended Splits

• Use sentence_id stratification to create train/val/test (e.g., 80/10/10) to avoid leakage across variants of the same sentence.
• When generating splits, keep all variants of a sentence in the same split.

Licensing

• Respect original SignTalk-GH and GSL licenses. If unsure, mark as "research-only" and require users to acknowledge source licenses.

Known Caveats

• Some GSL variants may be skipped if variant count exceeds alphabet length (A–Z).
• unmapped_videos.csv may exist if filenames don't match the expected pattern; review before release.
• Frame rates and resolutions are not normalized at this stage; preprocessing handles normalization.

Provenance & Reproducibility

• Notebook: testing-gsl-datasets.ipynb (A copy of the notebook is here)
• Random seed: 42
• Sampling knobs: TARGET_SAMPLE_COUNT, MIN_VARIANTS_PER_SENTENCE, MAX_VARIANTS_PER_SENTENCE, DOUBLE_SAMPLE

Changelog

• v1.0: Initial sampled release (~11k clips) with GSL integration.