📄 ArabicDialectHub: A Cross-Dialectal Arabic Learning Resource and Platform

Summary

ArabicDialectHub targets a gap in Arabic learning. Most apps teach only MSA, and dialect corpora such as MADAR or PADIC are built for NLP research, not for learners. The author contributes two things. The first is a collection of 552 phrases organized for teaching, spanning MSA along with five dialects: Lebanese, Syrian, Saudi, Emirati and Moroccan Darija. Darija is the default source variety. The phrases are grouped by theme and graded into three difficulty levels. Translations were drafted with Claude 3.5 and GPT-4, using prompts that described each dialect and the expected register. Drafts were then checked for cross-dialect consistency and reviewed by native speakers. The second contribution is an open-source web platform with four parts: a phrase browser, quizzes with automatically chosen distractors, synced progress tracking, and cultural notes. The paper reports no learning-outcome evaluation. Its claim of usefulness rests on a working system. Its value is mainly as open, extensible infrastructure for learning across dialects, especially for Darija speakers. The data and code are released under MIT.

Key points

• The resource has 552 phrases in six varieties, spread over 18 thematic categories and three difficulty tiers. It also includes 400 extra daily-conversation sentences.
• Each entry carries:
  • Arabic script and Latin romanization
  • a literal English gloss
  • usage notes on register and context for every dialect
• Validation was done by five native speakers, three for Darija and two for Lebanese. Syrian, Emirati and Saudi were never checked by native speakers, and no inter-annotator agreement was computed.
• Difficulty labels were assigned by Claude 3.5 without formal criteria. The paper includes no user studies and no audio.
• The stack is React 18/TypeScript, Clerk for authentication, Supabase/PostgreSQL with row-level security (tables phrases, phrase_progress and quiz_attempts), and Netlify deployment.

Interesting ideas

• I1. Candidate translations for all varieties were compared against each other to spot outliers, and validators picked or merged the best of several generations; agreement across parallel outputs works as a cheap filter before human review.
• I2. The word-ordering quiz has learners rebuild the target sentence from shuffled words, and the shuffle is forced to be far enough from the original that surface patterns can't give the answer away.
• I3. Multiple-choice distractors are drawn from the resource itself by lexical or phonological similarity, so wrong options are plausible rather than random.
• I4. The paper openly lists validation coverage per variety, separating checked items from merely generated ones.
• I5. Pragmatic knowledge is kept at two levels: short register and formality notes on each phrase, plus separate thematic cards such as expected greeting responses by region.

For Mova

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

• I1 (learn): Silver trees from witness parsers could be screened the same way, sending only outlier sentences to the LLM teacher tier instead of reviewing everything.
• I2 (back): Rebuilding word order from a meaning graph or bag of tokens is a natural test for linearization, and requiring a minimum scramble distance gives a built-in negative control against trivial copying.
• I3 (answer): School-test and fable-QA evaluations could build hard distractors from similar entities or phrases in the same text world, making tests check real understanding instead of elimination.
• I4 (learn): Each training or evaluation item should record who verified it (which parser, teacher tier or human), so metrics can be reported on verified-only subsets.
• I5 (frames): Speech-act handling could pair utterance-level register features with a reusable store of conventional exchange patterns, for example which response a greeting expects.