jamalinu/tarifit-catalan-public-services

🏔️ Tarifit–Catalan–Spanish Corpus for Public Services (TCSPS)

📌 Overview

TCSPS is a trilingual lexical and phraseological corpus covering Tarifit (Riffian Berber, Nador dialect) paired with Catalan and Spanish translations. It is designed to support natural language processing research and language technology development for one of the most underrepresented Berber varieties in the digital world.

Tarifit (ISO 639-3: rif) is spoken by approximately 1.2–4 million people in the Rif region of northern Morocco (provinces of Nador, Al Hoceima, Driouch) and by a significant diaspora community in the Netherlands, Belgium, Spain, France and Germany. Despite this speaker base, it remains severely underrepresented in NLP datasets and language models.

This corpus is especially relevant for Catalan–Tarifit language contact contexts, given the large Riffian community residing in Catalonia and other Catalan-speaking territories.

⚠️ Validation note: Translations marked "Revisar traducción" in the translation_es field are pending native speaker review. All Catalan translations have been validated by a native speaker with professional linguistic training.

---

📊 Dataset Statistics (v0.7)

Field	Value
Total entries	394
Thematic domains	15
Dialect	Nador (Central Riffian)
Scripts	Latin transcription (all), Tifinagh (4 entries)
Languages	Tarifit (rif), Catalan (ca), Spanish (es)
Source	University of Barcelona multilingual guide (ub_guia)

Entry types

Type	Count
phrase	125
sentence	118
vocabulary	91
word	60

Thematic domains

Domain	Label	Count
universitat	3	123
restauracio	13	53
salut	5	38
comerc	7	24
transports	12	23
esports	9	22
comunicacio_basica	1	21
telefon	11	17
lleure	14	17
allotjament	4	12
banc	8	11
correus	10	11
saludos	0	10
biblioteca	6	9
orientacio	2	3

---

📁 Dataset Structure

Each entry contains the following fields:

Field	Type	Description
id	string	Unique identifier (rif_0001 … rif_0394)
text_rif	string	Tarifit text in Latin transcription (Nador standard)
tifinagh	string	Tarifit in Tifinagh script (partially filled)
translation_cat	string	Catalan translation (validated)
translation_es	string	Spanish translation (partially pending review)
domain	string	Thematic domain (15 categories)
label	int	Numeric domain ID (0–14)
type	string	Entry type: word, vocabulary, phrase, sentence
subtopic	string	Sub-category within domain (optional)
source	string	Data source identifier
dialect_var	string	Dialect variant (always nador)

Example entry

{
  "id": "rif_0001",
  "text_rif": "Azul",
  "tifinagh": "ⴰⵣⵓⵍ",
  "translation_cat": "Hola / Bon dia / Bona tarda",
  "translation_es": "Hola / Buenos días",
  "source": "ub_guia",
  "dialect_var": "nador",
  "domain": "saludos",
  "label": 0,
  "type": "sentence",
  "subtopic": ""
}

---

🛠️ Intended Uses

• Machine translation (MT): Tarifit ↔ Catalan, Tarifit ↔ Spanish low-resource MT experiments
• Domain classification: Multi-class text classification in Tarifit (15 domains)
• Lexical resources: Tarifit vocabulary with multilingual glosses
• Low-resource NLP benchmarking: Few-shot and zero-shot evaluation for Tarifit
• Public service NLP: Dialogue systems for Tarifit-speaking communities accessing Catalan public services

Out-of-scope uses

This dataset should not be used to build surveillance systems or any application that could harm Tarifit-speaking communities.

---

🔗 Related Resources

• Space: jamalinu/tarifit-assistant — a dialogue assistant built on this corpus
• Tamazight-NLP community: huggingface.co/Tamazight-NLP — broader Amazigh NLP ecosystem
• Common Voice Tamazight: Mozilla Common Voice speech data for Tamazight
• NLLB-200: Meta's multilingual MT model (includes zgh Standard Moroccan Tamazight, adjacent to Tarifit)

---

⚠️ Limitations

• Single source: All entries derive from one University of Barcelona multilingual guide, which limits domain and register diversity.
• Single dialect: Only the Nador (Central Riffian) dialect is represented. Al Hoceima, Driouch and Berkane varieties are not covered.
• Tifinagh incomplete: Only 4 of 394 entries have Tifinagh script filled in.
• Spanish pending review: Many translation_es fields contain placeholder text (Revisar traducción) and require expert validation.
• Small size: 394 entries is sufficient for lexical reference and classification experiments but insufficient for training MT models from scratch.

---

📜 Citation

If you use this dataset in your research, please cite it as:

@dataset{tarifit_catalan_public_services_2025,
  author    = {Jamal Saghraoui. (jamalinu)},
  title     = {Tarifit–Catalan–Spanish Corpus for Public Services (TCSPS)},
  year      = {2025},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/jamalinu/tarifit-catalan-public-services},
  license   = {CC-BY-4.0}
}

---

🗺️ Version History

Version	Entries	Changes
v0.7	394	Initial public release. Tarifit–Catalan validated. Spanish partially pending.

---

📬 Contact

For corrections, contributions or collaboration proposals — especially from native Tarifit speakers — please open a discussion in the Community tab.