Datasetpublic

Alhasan/egyptian-nlu

Egyptian Arabic voice-assistant NLU — dataset Egyptian Arabic commands paired with intent + slot annotations, in the schema of Amazon MASSIVE (60 intents, 55 slot types). Two parts, and the difference matters: File Rows Origin egy_test.jsonl 200 Written and annotated by hand by a native Egyptian speaker — the benchmark egy_synth_train.jsonl 6,509 LLM-generated Egyptian rewrites of MASSIVE ar-SA training items egy_synth_dev.jsonl 730 Same, held out by seed… See the full description on the dataset page: https://huggingface.co/datasets/Alhasan/egyptian-nlu.

sourceHugging Facecc-by-4.0updated 22h agoView on Hugging Face

0likes60downloads

Dataset Card

Egyptian Arabic voice-assistant NLU — dataset

Egyptian Arabic commands paired with intent + slot annotations, in the schema of Amazon MASSIVE (60 intents, 55 slot types).

Two parts, and the difference matters:

File	Rows	Origin
`egy_test.jsonl`	200	Written and annotated by hand by a native Egyptian speaker — the benchmark
`egy_synth_train.jsonl`	6,509	LLM-generated Egyptian rewrites of MASSIVE `ar-SA` training items
`egy_synth_dev.jsonl`	730	Same, held out by seed

egy_test.csv is the raw annotation sheet, with each item's Saudi/MSA source next to the Egyptian rewrite.

Format

json

{"id": "egy-main-massive-3161", "utt": "صحيني بكرة الساعة ستة الصبح",
 "intent": "alarm_set",
 "slots": [{"type": "date", "value": "بكرة", "start": 6, "end": 10},
           {"type": "time", "value": "ستة الصبح", "start": 17, "end": 26}],
 "variety": "egy", "source": "human", "split": "test", "scenario": "alarm"}

Slot values are verbatim spans with character offsets, following MASSIVE's conventions (attached prepositions stay inside the span: لستة الصبح).

How the synthetic data was made

MASSIVE ar-SA training items (mostly Saudi colloquial and MSA) were rewritten into Egyptian by Claude Sonnet 5 through the Batches API, two variants per seed, in MASSIVE's bracket notation so slot spans stay exact. 8,002 candidates → 7,239 kept by rule-based filters — no LLM judge:

Rejected because	Count
Too close to the seed (not really rewritten)	428
Exact duplicate	220
Near duplicate	60
Slot types changed	55

Known limitations

—The generated style is not the natural one. Egyptian marker words appear in 60% of generated sentences but 32% of the human test set; English code-switching in 27% versus 10%. The generator writes "more Egyptian" than Egyptians do. Validate against the human split before trusting it.
—19 of the 200 test items also appear verbatim in training data — short commands like "امسح المنبه" that any two writers phrase identically. Scores on the clean 181 are reported alongside in the project's results.
—Domain is bounded by MASSIVE: smart-speaker commands, no long-form or multi-turn speech.

Provenance and licence

Derived from Amazon MASSIVE (CC BY 4.0) and redistributed under CC BY 4.0 with attribution.

bibtex

@inproceedings{fitzgerald-etal-2023-massive,
  title = {{MASSIVE}: A 1{M}-Example Multilingual Natural Language Understanding Dataset},
  author = {FitzGerald, Jack and others},
  booktitle = {Proceedings of ACL},
  year = {2023}
}

Code, evaluation harness and full results: [GitHub repo](https://github.com/Alhasan-Abdellatif/Egyptian-Arabic-understanding-for-voice-assistants)