📄 Lahjawi: Arabic Cross-Dialect Translator

Summary

Arabic speakers mostly communicate in regional dialects, and these dialects differ in vocabulary, word order, idioms and morphology. Earlier machine translation work usually mapped one dialect, or a few, into Modern Standard Arabic (MSA). Direct dialect-to-dialect translation has received little attention. The authors introduce Lahjawi, a family of models fine-tuned from Kuwain-1.5B, their own small Arabic language model.

• Lahjawi-D2D translates directly among 15 dialects.
• Lahjawi-D2MSA normalizes any dialect into MSA.

The training data comes from seven open parallel corpora, cast into two prompt templates. For the dialect-to-dialect set, every possible pairing of aligned dialect sentences was generated. Evaluation used NADI-2024 and the MADAR test set, plus human judgments.

BLEU is modest: about 9.6 for D2MSA and 9.9 for D2D. Human raters judged 58% of outputs accurate and 78% fluent. Results are strongest for Levantine dialects and weakest for Maghrebi ones. The paper matters mainly as the first broad cross-dialect Arabic translator and as a map of where data imbalance and dialect distance hurt.

Key points

• Training data. Sources are MADAR, PADIC, NADI (2023), Dial2MSA, Arabic STS, UFAL Parallel Corpus of North Levantine 1.0 and MDPCA, normalized for characters, numerals, punctuation and spacing.
• D2MSA set. It has 197,042 samples and is heavily skewed, with Syrian making up 66% and other dialects between 0.8% and 5.7%.
• D2D set. It has 266,871 samples spanning 210 directed pairs over 15 dialects, and Levantine and Maghrebi pairs are over-represented.
• Evaluation and comparison. Testing used NADI-2024 DA-MSA (Egyptian, Emirati, Jordanian, Palestinian) and the MADAR test set. Lahjawi-D2MSA trails shared-task systems built on Jais-13B or LLaMA-8B, or on augmented data.
• Findings.
  • A Palestinian-trained model scored 19.34 BLEU on Jordanian data, while the Jordanian-trained model scored only 5.67 on its own dialect.
  • Translation is direction-dependent: Qatari→Iraqi reached 17.21 BLEU, but Iraqi→Qatari only 6.02.

Interesting ideas

• I1. Building all ordered pairs from sentences aligned through a shared pivot turns a multi-way parallel corpus into training data for every direction at once.
• I2. A single model serves many directions, with the target variety named explicitly inside a fixed prompt template, instead of separate models per pair.
• I3. Cross-testing models across neighbouring varieties showed that a small dataset, not dialect difficulty, explained a weak result, because pooling similar varieties helped.
• I4. Scores were reported per direction and revealed strong asymmetry, so averaging both directions would hide real failures.
• I5. Weak results were explained by checking which varieties the base model likely saw during pre-training, not only fine-tuning data counts.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (back): Sentences aligned across several English registers or paraphrase sources could be expanded into all-direction pairs to test text → graph → text style transfer through the meta-world.
• I2 (learn): LLM teacher tiers could use one templated prompt with an explicit target-schema slot (e.g., UD version or frame inventory), keeping silver data uniform and comparable.
• I3 (tree): When gparse underperforms on a small English treebank, train on one treebank and test on another, then pool related treebanks, to separate data scarcity from genuine domain mismatch.
• I4 (tree): UD dialect conversion rules should be evaluated A→B and B→A separately, with round-trip checks, to expose lossy conversion directions.
• I5 (learn): Evaluation reports could log the base model's pre-training exposure per genre or source, so gains are not wrongly credited to Mova's own training.