📄 ArzEn-MultiGenre: An aligned parallel dataset of Egyptian Arabic song lyrics, novels, and subtitles, with English translations

Summary

Parallel resources pairing Egyptian Arabic with English are scarce and small. Most are built from conversational or web genres such as SMS, forums and telephone speech, and many rely on crowdsourced translation. This data article presents ArzEn-MultiGenre, a human-made gold-standard corpus of 25,557 Egyptian Arabic–English segment pairs in three genres absent from earlier resources: song lyrics, novels and Netflix subtitles.

Each genre was built differently:

• Songs: lyrics were crawled and then translated from scratch by professional subtitlers following explicit guidelines.
• Novels: published novels and their Egyptian Arabic translations were scanned, converted with OCR and proofread, keeping the original spelling.
• Subtitles: these were transcribed by hand.

Annotators then segmented and aligned everything using a fixed set of correspondence types. They cross-reviewed each other's work, and the author adjudicated.

As a usage demonstration, the author customized Google Translate with AutoML Translation on the data, which improved BLEU in every genre. The corpus is useful for benchmarking MT on dialectal, idiomatic and culture-heavy text, for few-shot LLM adaptation, and for translation studies, for example on how gender is rendered.

Key points

• Size and sources: 25,557 segment pairs in total.

  • Novels: 5,226 pairs from The Old Man and the Sea, The Little Prince and The Stranger.
  • Subtitles: 17,265 pairs from 12 episodes of Paranormal and Finding Ola.
  • Songs: 3,066 pairs from 280 lyrics.

• Distribution: released as Excel bilingual tables on Mendeley Data (10.17632/6k97jty9xg.4).

• AutoML customization of Google NMT raised BLEU:

  Genre	Before	After
  Songs	8.95	11.87
  Novels	11.84	17.28
  Subtitles	12.54	19.05

• Alignment types: one-to-one, one-to-many, many-to-one, source-without-target and target-without-source. Zero correspondences appear 483 times in subtitles, 32 times in novels and never in songs.

• Caveats:

  • Google's tooling silently drops pairs that share a source segment, so the AutoML split counts are smaller than the corpus counts.
  • The Arabic versions of two novels were translated from French, while the English sides are separate translations.

Interesting ideas

• I1. Alignment uses an explicit set of correspondence types that includes "target with no source," which captures cases where the translator added an explanation of an idiom or a cultural reference; this turns translator additions into labeled data rather than noise.
• I2. Proofreaders were told to keep the original, non-standard dialect spelling and punctuation instead of normalizing it, which preserves real-world variation for downstream study.
• I3. Statistics are broken down by genre and by segment length (1 word, 2–5 words, 6+ words), exposing how many one-word segments, such as discourse fillers, each genre contains.
• I4. Quality control combined a peer swap, in which annotators reviewed each other's portions, with a final single adjudicator. This is a cheap layered agreement scheme.
• I5. The paper flags that an external tool deduplicated pairs by source segment and so changed the evaluation counts, a hidden preprocessing effect worth reporting.

For Mova

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

• I1 (back): when translating through the meta-world, mark spans that render implicit cultural knowledge as explicit text, so generated explanations stay traceable to the graph node that licensed them.
• I2 (tokens): store raw surface forms untouched and put normalization in a separate, reversible layer, so lemma decisions can always be explained against the original string.
• I3 (learn): stratify evaluation by genre and length bucket so one-word utterances like fillers and interjections are scored separately and do not hide failures on full sentences.
• I4 (learn): mirror this for silver data by letting witness parsers cross-check each other's outputs and sending only disagreements to an LLM-teacher adjudicator tier.
• I5 (learn): log every dedup or filter step an external toolkit applies, so that evaluation numbers remain comparable and explainable across runs.