OmarSY11/ferman-nlu

Ferman NLU: intent and slot data for English, Arabic, Sorani and Badini

Code, app and baseline results: https://github.com/omarGH99/ferman-ai-assistant

Load it with:

from datasets import load_dataset
ds = load_dataset("OmarSY11/ferman-nlu", "v2")["train"]   # one table; filter on `partition`
test = ds.filter(lambda r: r["partition"] == "test")

The partition column holds train, validation and test. Hugging Face shows all rows as a single split, so filter on that column.

Intent classification and slot filling data for voice/text assistant commands in English, Arabic, Sorani Kurdish and Badini Kurdish (Arabic script).

Files

File	Rows	Notes
XLMR_dataset_4lang_v2.csv	27,445	Cleaned release. Trained and evaluated the deployed v2 models.
XLMR_dataset_4lang_v3.csv	31,829	v2 plus machine-translated Kurdish rows (see below). No model has been trained or evaluated on v3 yet.

Columns: language (en/ar/sorani/badini), partition (train/validation/test), intent, utt (the utterance), annot (the utterance with inline slots, [slot : value]; empty when there are no slot annotations).

Composition (v2)

Language	train	validation	test	rows with slot annotation
English	7,753	1,390	1,954	11,097
Arabic	7,753	1,390	1,954	304
Sorani	2,065	191	370	2,626
Badini	2,064	191	370	2,625

34 intents across alarms, calendar, lists, news, weather, QA, email, transport, takeaway, recommendations, cooking, date/time and general chat. Per-intent counts are in dataset_v2_report.txt, which also records the 26 intents and 15,096 rows dropped from v1 and why.

How it was built

• Source: the English and Arabic rows (and the intent label set, slot syntax and train/validation/test split) come from MASSIVE (FitzGerald et al., 2022; Amazon; CC BY 4.0). This dataset is a modified derivative: intents out of scope for this app (smart-home, music, volume) were removed, junk rows dropped and spelling fixed (see dataset_v2_report.txt).
• Sorani and Badini are not in MASSIVE. They were added by the author using machine translation of the English rows. This applies to the Kurdish rows in v2 as well as v3. The author, a native Badini speaker and a non-native Sorani speaker, informally reviewed the Sorani and Badini rows and judged them good. This is not an independent verification: there was no second annotator, no agreement measurement, and the Sorani review was by a non-native speaker.
• v3 augmentation (ml/augment_kurdish.py): Kurdish intents were very unevenly covered (e.g. calendar_set had 806 English rows but 19 Sorani). v3 adds more NLLB-200 translations (facebook/nllb-200-distilled-600M): English → Sorani (ckb_Arab) and English → Kurmanji (kmr_Latn), the latter transliterated into Arabic script by ml/kurmanji_translit.py to stand in for Badini. Translations are round-tripped back to English and dropped if they drift too far; slot values must be found in the translated sentence or the row keeps its intent label without annotation. Added rows are only in train (+2,249 Badini, +2,135 Sorani); validation and test are unchanged from v2.

Known limitations

• Machine-translated Kurdish has only an informal review by its author (native Badini, non-native Sorani); a native Sorani speaker has not checked it. It may contain unnatural phrasing, and the Kurmanji→Arabic-script conversion is a stand-in for real Badini, not a native-speaker rendering. There is no marker column distinguishing translated from original rows; they are the rows in v3 that are not in v2.
• Small Kurdish evaluation sets: 370 test rows per dialect (vs 1,954 for English/Arabic), so Kurdish scores have wide confidence intervals.
• Arabic slots are essentially absent: 304 annotated rows, all in train, none in test, so Arabic slot extraction cannot be evaluated with this data. MASSIVE itself provides human slot labels for Arabic (ar-SA); they were not carried over into this release. All of this release's Arabic utterances appear in MASSIVE ar-SA, so the labels can be restored from the source.
• Class imbalance: intents range from 24 (cooking_query) to 2,497 (weather_query) rows, and the Kurdish distribution is more skewed than English.
• Dialect coverage: Sorani and Badini only. Other varieties (Kurmanji in Latin script, Gorani, Iraqi Arabic dialect variation beyond what the Arabic rows contain) are not covered.
• Text is not personal data, but data collected from app users (via ml/export_dataset.py) is opt-in and is not included in these files; neither CSV contains any user-collected data.

Baseline results

Measured on the v2 test split (full output: eval_results_v2.txt):

Language	Intent accuracy	Intent macro-F1	Slot F1
English	0.887	0.847	0.816
Arabic	0.823	0.761	n/a
Badini	0.719	0.682	0.623
Sorani	0.697	0.647	0.643

License and citation

CC BY 4.0, matching MASSIVE's license. If you use this dataset you must also credit MASSIVE. Changes from the original are described above.

@misc{fitzgerald2022massive,
  title  = {MASSIVE: A 1M-Example Multilingual Natural Language Understanding
            Dataset with 51 Typologically-Diverse Languages},
  author = {FitzGerald, Jack and others},
  year   = {2022},
  eprint = {2204.08582},
  archivePrefix = {arXiv}
}

Please also cite the project repository: https://github.com/omarGH99/ferman-ai-assistant