rabeeeehh/arabizi-kit-corpus

ArabiziKit Corpus

LLM-annotated sentences of real social-media Arabizi (Arabic written in Latin letters and digits) with Arabic-script references and dialect tags, stratified into train / dev / test splits.

• 355 annotated sentences harvested from public Hugging Face datasets (filtered for Arabizi, URLs/mentions/hashtags/emoji stripped, split into sentences)
• 323 usable rows after dropping empty references: train 226 / dev 48 / test 49, stratified by dialect
• Inter-annotator agreement 0.171 exact on normalized Arabic over a deterministic 10% double-annotated sample — the honest measure of how ambiguous Arabizi rendering is
• Produced by the ArabiziKit corpus pipeline (https://github.com/rb2625/arabizi-kit): arabizikit corpus run

Each split row contains the Arabizi source, the Arabic-script reference, and a dialect tag (MSA / Egyptian / Levantine / Maghrebi / Gulf).

Files

file	contents
annotated.jsonl	all 355 annotated sentences with references and dialect tags
train.json / dev.json / test.json	stratified splits in benchmark format

Evaluation sets

The paper also evaluates on three external parallel Arabizi/Arabic datasets with gold references, imported from Hugging Face and not redistributed here — rebuild them in one command each:

arabizikit corpus import-hf --dataset arbml/Arabizi_Transliteration --arabizi-field Arabize --arabic-field Arabic
arabizikit corpus import-hf --dataset akhanafer/arabic-to-arabizi --arabizi-field arabizi --arabic-field arabic
arabizikit corpus import-hf --dataset elkababi2/Darija-Text-Ar-Arabizi --arabizi-field darija_Latn --arabic-field darija_Arab_new

Usage

pip install arabizikit
arabizikit eval --data train.json   # score the rules on any split
arabizikit eval --data test.json --model   # learned layer

Paper

"ArabiziKit: An Open, Hybrid, Benchmark-Driven Arabizi to Arabic-Script Transliteration System" (arXiv, cs.CL) — see https://github.com/rb2625/arabizi-kit for the library, demo, and benchmark.