ArzEn-MultiGenre: Egyptian Arabic–English Corpus
- ArzEn-MultiGenre is a manually aligned Egyptian Arabic–English parallel corpus covering song lyrics, novels, and TV show subtitles with a focus on dialectal variety.
- It preserves authentic non-standard orthography and code-mixing, providing a realistic linguistic resource for machine translation and cross-linguistic analysis.
- The dataset supports domain-specific MT benchmarking, few-shot large language model adaptation, and translation studies, backed by expert human translations.
ArzEn-MultiGenre is a parallel Egyptian Arabic–English dataset composed of manually translated and manually aligned segment pairs drawn from three genres—song lyrics, novels, and TV show subtitles—and designed to expand the coverage of dialectal Arabic–English resources beyond the Modern Standard Arabic-centered corpora that dominate the literature (Al-Sabbagh, 2 Aug 2025). Its central contribution is not merely scale, but the combination of genre novelty and gold-standard human production: the dataset introduces genres described as absent from prior Egyptian Arabic–English parallel resources, and its translation and alignment were carried out by human experts rather than by crowdsourcing or fully automatic pipelines (Al-Sabbagh, 2 Aug 2025). The resulting corpus contains 25,557 aligned segment pairs and is positioned simultaneously as a machine translation benchmark, a resource for few-shot LLM adaptation, a translation-studies dataset, and a practical translation memory for pedagogy and professional use (Al-Sabbagh, 2 Aug 2025).
1. Definition, language setting, and corpus identity
ArzEn-MultiGenre is explicitly an Egyptian Arabic ↔ English parallel resource, with the Arabic side written in Arabic script and characterized as dialectal Egyptian Arabic rather than Modern Standard Arabic (Al-Sabbagh, 2 Aug 2025). The paper repeatedly contrasts this design choice with earlier Arabic–English corpora whose Arabic component is predominantly MSA, framing ArzEn-MultiGenre as an intervention in the relative scarcity of dialectal Arabic–English parallel data (Al-Sabbagh, 2 Aug 2025).
A defining property of the Arabic side is the deliberate preservation of source orthography. Egyptian Arabic is described as lacking fixed spelling and punctuation rules, and the compilers instructed annotators and proofreaders not to normalize the source materials for the novels and subtitles (Al-Sabbagh, 2 Aug 2025). This implies that the corpus preserves orthographic variation as an intrinsic feature rather than treating it as noise. In subtitles, the treatment of code-mixing is constrained but not eliminated: sentence-level English insertions were excluded from transcription, whereas word-level code-mixing could remain, either in Arabic script for commonly Arabized items or in mixed Arabic/English writing when the English word is not typically Arabized in script (Al-Sabbagh, 2 Aug 2025).
The designation “MultiGenre” refers specifically to three genres: song lyrics, novels, and TV show subtitles (Al-Sabbagh, 2 Aug 2025). This is a genre expansion within the Egyptian Arabic–English parallel-corpus literature rather than a claim to exhaustiveness over Egyptian registers. The resource is organized into three folders—songs, novels, and subtitles—and the genres are the primary organizing principle of the corpus release (Al-Sabbagh, 2 Aug 2025).
2. Composition, segmentation, and quantitative profile
The corpus contains 25,557 aligned segment pairs in total (Al-Sabbagh, 2 Aug 2025). Its genre distribution is asymmetric: subtitles are the largest component with 11,882 segment pairs, novels contribute 7,809, and songs contribute 5,866 (Al-Sabbagh, 2 Aug 2025).
| Genre | Segment pairs | Distinctive source basis |
|---|---|---|
| Subtitles | 11,882 | Two Netflix series |
| Novels | 7,809 | Three literary works |
| Songs | 5,866 | 280 Amr Diab lyrics |
The lexical statistics reported in Table 4 underscore genre heterogeneity. Novels contain 56,527 Arz word tokens and 78,890 English word tokens, with 15,590 Arz word types and 10,246 English word types; subtitles contain 80,273 Arz tokens and 99,884 English tokens, with 12,798 and 10,032 types respectively; songs contain 17,858 Arz tokens and 31,294 English tokens, with 5,791 and 3,853 types (Al-Sabbagh, 2 Aug 2025). The total counts are 154,658 Arz word tokens, 210,068 English word tokens, 29,179 Arz word types, and 18,131 English word types, with type-token ratios of 19% for Arz and 9% for English (Al-Sabbagh, 2 Aug 2025).
The operational unit is the segment. Segmentation was defined so that each segment ends with either a major punctuation mark—period, exclamation mark, or question mark—or a line break (Al-Sabbagh, 2 Aug 2025). As a result, segments may be one-word discourse markers, stand-alone phrases, simple sentences, complex sentences, or compound sentences (Al-Sabbagh, 2 Aug 2025). The paper highlights that this segmentation regime yields challenging material containing one-word segments, idiomatic expressions, culture-specific references, and slang words (Al-Sabbagh, 2 Aug 2025).
The corpus also records genre-specific irregularities in alignment. The paper states that subtitles contain 483 segments with zero correspondences, novels contain 32, and songs contain none (Al-Sabbagh, 2 Aug 2025). The absence of zero correspondences in songs is attributed to the instruction that all song segments had to be translated (Al-Sabbagh, 2 Aug 2025).
A further structural specificity lies in the internal content selection of each genre. The songs folder contains one file with all lyrics together, covering 280 song lyrics by Amr Diab from 1983 to 2023, with metadata columns for song name, year of album release, album name, song number in the dataset, Egyptian Arabic lyrics, and English translation (Al-Sabbagh, 2 Aug 2025). The novels folder contains three files, one per novel: The Old Man and the Sea, The Little Prince, and The Stranger (Al-Sabbagh, 2 Aug 2025). The subtitles folder contains 12 files, one per episode, from the Netflix series Paranormal and Finding Ola (Al-Sabbagh, 2 Aug 2025).
3. Data creation pipeline and annotation methodology
The corpus was built through genre-specific acquisition pipelines, but the common design principle is strong human control over both bilingual transfer and alignment (Al-Sabbagh, 2 Aug 2025). This is the basis for the paper’s description of ArzEn-MultiGenre as a gold-standard dataset (Al-Sabbagh, 2 Aug 2025).
For songs, the Arabic lyrics were collected from Aghani Lyrics / Alaghani Lyrics using an in-house web crawler, then parsed and cleaned with BeautifulSoup4 v4.12.2, after which duplicate lines were removed (Al-Sabbagh, 2 Aug 2025). The English translations were produced manually by four translators and checked by one reviewer (Al-Sabbagh, 2 Aug 2025). All translators and the reviewer were native speakers of Egyptian Arabic; the translators had four years of experience in Egyptian Arabic–English subtitling for Netflix, and the reviewer had six years (Al-Sabbagh, 2 Aug 2025). The corpus was divided equally among the four translators, and the reviewer checked translations for accuracy and consistency (Al-Sabbagh, 2 Aug 2025). The guidelines explicitly prohibited machine translation and AI use: “Translators must not use machine translation or artificial intelligence applications; sentences must be translated manually from scratch” (Al-Sabbagh, 2 Aug 2025).
For novels, the Egyptian Arabic versions were purchased from Hunna and the English versions from Amazon UK (Al-Sabbagh, 2 Aug 2025). Because both sides were only available as hard copies, they were scanned and converted into text using Sotoor, an OCR system supporting Arabic and batch API processing (Al-Sabbagh, 2 Aug 2025). Proofreaders then manually checked the OCR output. These proofreaders were native speakers of Egyptian Arabic and graduates of English language departments in Egyptian universities, and they were instructed to preserve original spelling and punctuation rather than normalize the texts (Al-Sabbagh, 2 Aug 2025). The Arabic translations are attributed to named professional translators: Magdy Abdelhadi translated The Old Man and the Sea, while Hector Fahmy translated The Little Prince and The Stranger; both are described as certified translators in Egypt working for Hunna (Al-Sabbagh, 2 Aug 2025).
For subtitles, the source material came from Netflix subtitle streams for Paranormal and Finding Ola (Al-Sabbagh, 2 Aug 2025). The paper notes that Netflix offers Egyptian Arabic subtitles mainly in SDH form, and English subtitles in both SDH and non-SDH forms, with potential differences not only in sound-effect information but in dialogue translation itself (Al-Sabbagh, 2 Aug 2025). Arabic transcribers fluent in English were hired to manually transcribe subtitle text files according to explicit guidelines (Al-Sabbagh, 2 Aug 2025). These guidelines required exclusion of non-dialogue information from Egyptian Arabic subtitles, transcription of both English SDH and non-SDH while excluding sound and music descriptions, preservation of on-screen Egyptian Arabic spelling and punctuation, exclusion of sentence-level English code-mixing, and controlled treatment of word-level code-mixing (Al-Sabbagh, 2 Aug 2025).
Alignment was then performed manually in Microsoft Excel by four annotators who were native Arabic speakers, well acquainted with Egyptian Arabic, and graduates of English departments in Egypt (Al-Sabbagh, 2 Aug 2025). The allowed alignment correspondences were defined as one-to-one, one-to-many, many-to-one, one/many-to-zero, and zero-to-one/many (Al-Sabbagh, 2 Aug 2025). Quality control proceeded in stages: the corpus was equally divided among the four annotators, annotators exchanged parts to review one another’s work after completion, and the author reviewed all alignments as adjudicator (Al-Sabbagh, 2 Aug 2025).
4. Translation principles, textual fidelity, and linguistic characteristics
The dataset’s translation regime is intentionally non-literal where literalness would distort meaning. For song translation, the guidelines instructed translators to aim for natural translations reflecting intended meaning, avoid literal translation of idioms, figurative language, and culture-specific references, use mainstream English without slang, use American English spelling, ignore singability and rhyme, preserve punctuation in the same relative position, and listen to songs to resolve ambiguous words (Al-Sabbagh, 2 Aug 2025). These instructions imply that the English side is optimized for semantic and pragmatic adequacy rather than formal mimicry.
The Arabic side is equally distinctive. It is dialectal Egyptian Arabic in Arabic script, not transliterated into Latin characters, and not normalized to a standardized orthography (Al-Sabbagh, 2 Aug 2025). This preservationist stance gives the corpus value for lexical semantics and sociolinguistically grounded translation analysis, but it also introduces preprocessing challenges for models sensitive to spelling variation and punctuation inconsistency (Al-Sabbagh, 2 Aug 2025). A plausible implication is that benchmarking on ArzEn-MultiGenre probes not only bilingual transfer quality but also robustness to non-standard orthographic variation.
The genres encode different linguistic pressures. Songs foreground figurative language and culture-specific references; subtitles foreground short turns, discourse markers, slang, and platform-specific subtitle conventions; novels contribute longer literary segments and translator-mediated stylistic variation (Al-Sabbagh, 2 Aug 2025). The paper explicitly emphasizes one-word segments, idiomatic expressions, cultural references, and slang as central sources of difficulty for automated systems (Al-Sabbagh, 2 Aug 2025).
Two of the novels also raise a strict parallelism issue. The Little Prince and The Stranger were originally written in French; the English side uses published English translations by Katherine Woods and Stuart Gilbert, while the Egyptian Arabic versions by Hector Fahmy were translated directly from French, not from English (Al-Sabbagh, 2 Aug 2025). Thus the English–Egyptian Arabic pairing for these two books is not a direct bilingual translation relation in the narrowest sense, but rather a parallel relation between translations derived from a common source language (Al-Sabbagh, 2 Aug 2025). This is a critical caution for work on fine-grained translation equivalence or alignment-based semantic correspondence.
5. Benchmarking value and demonstrated uses
A primary intended use of ArzEn-MultiGenre is benchmarking machine translation across genres (Al-Sabbagh, 2 Aug 2025). The paper argues that the corpus is challenging because of one-word segments, idioms, cultural references, slang, and fast-changing dialect vocabulary and semantics (Al-Sabbagh, 2 Aug 2025). It also states that the dataset can be used to fine-tune LLMs, especially in few-shot settings, and to adapt commercial MT systems such as Google Translate (Al-Sabbagh, 2 Aug 2025).
The paper provides one explicit adaptation result through AutoML Translation. Customizing Google Translate with ArzEn-MultiGenre improved BLEU in all three genres: songs increased from 8.95 to 11.87, novels from 11.84 to 17.28, and subtitles from 12.54 to 19.05 (Al-Sabbagh, 2 Aug 2025). These results are presented as evidence that the corpus supports domain- and genre-specific MT customization.
Beyond MT engineering, the dataset is explicitly positioned for translation studies, cross-linguistic analysis, and lexical semantics (Al-Sabbagh, 2 Aug 2025). The paper gives examples such as comparing translation strategies across genres, studying the translation of figurative language and culture-specific references, and investigating how gender is translated in song lyrics (Al-Sabbagh, 2 Aug 2025). The gender question is singled out because Arabic is heavily gendered, Egyptian Arabic often uses masculine forms for female lovers, and gender remains a significant challenge for Arabic MT (Al-Sabbagh, 2 Aug 2025).
The corpus is also framed as useful for pedagogy and professional translation practice (Al-Sabbagh, 2 Aug 2025). It is proposed as a training resource for translation students and as a translation memory for professional translators, with explicit reference to contemporary demand from streaming platforms subtitling Egyptian Arabic audiovisual content into English and the increased translation of world literature into Egyptian Arabic rather than MSA (Al-Sabbagh, 2 Aug 2025).
Its relevance extends to adjacent code-switched Egyptian Arabic–English modeling. Although ArzEn-MultiGenre is not itself a speech corpus, "ArzEn-LLM: Code-Switched Egyptian Arabic-English Translation and Speech Recognition Using LLMs" reports using the entire parallel corpora presented in ArzEn-MultiGenre as extra data for model training over ArzEn-ST, treating its 25,557 manually aligned segment pairs from song lyrics, novels, and TV show subtitles as auxiliary parallel text (Heakl et al., 2024). That paper also notes that adding such extra genre-diverse data can help smaller models more than larger ones, while not uniformly benefiting the strongest model (Heakl et al., 2024). This suggests that ArzEn-MultiGenre is not only a benchmark corpus but also a resource for cross-domain transfer in Egyptian Arabic–English systems.
6. Novelty relative to prior resources and relation to other Arabic datasets
The novelty claim of ArzEn-MultiGenre is twofold: newly introduced genre coverage and expert-produced gold-standard quality (Al-Sabbagh, 2 Aug 2025). The paper contrasts it with earlier Egyptian Arabic–English resources focused on personal interviews, weblogs, SMS/chat, discussion forums, travel questions and answers, and unscripted telephone conversations (Al-Sabbagh, 2 Aug 2025). In that comparison, songs, novels, and subtitles are presented as genres not represented in current parallel Egyptian Arabic–English datasets (Al-Sabbagh, 2 Aug 2025).
The paper also distinguishes ArzEn-MultiGenre from two nearby resources. The Habibi corpus contains song lyrics from multiple Arabic dialects including Egyptian Arabic, but it is monolingual and therefore does not provide Egyptian Arabic–English parallel data (Al-Sabbagh, 2 Aug 2025). OPUS subtitles contains many English–Arabic subtitle pairs, but its Arabic side is Modern Standard Arabic rather than Egyptian Arabic (Al-Sabbagh, 2 Aug 2025). These contrasts define the niche occupied by ArzEn-MultiGenre: dialectal Egyptian Arabic, bilingual, and genre-expanded.
In size, the paper reports the following comparison values: ArzEn-MultiGenre contains 25,557 segment pairs, compared with 2,600 for Zbib et al. [3], 12,000 for Bouamor et al. [6], and 35,892 for Kumar et al. 8. The paper explicitly notes that Kumar et al. is larger, but its translations were produced by crowdsourced amateur translators, whereas ArzEn-MultiGenre relies on professional translators and expert annotators (Al-Sabbagh, 2 Aug 2025). The significance of the resource therefore lies less in maximum segment count than in the combination of dialect specificity, genre range, and annotation provenance.
A common misconception is to interpret “ArzEn-MultiGenre” as a broad multilingual or multi-register Arabic benchmark. The dataset is neither multilingual in the sense of spanning multiple target languages nor general across all Egyptian Arabic textual practices (Al-Sabbagh, 2 Aug 2025). It is specifically Egyptian Arabic–English, and its “multigenre” status is confined to three curated genres. By contrast, a dataset such as ALHD is multigenre in a different sense—news, social media, and reviews—and is Arabic-only, designed for human-versus-LLM text detection rather than bilingual alignment (Khairallah et al., 3 Oct 2025). The comparison clarifies that “multigenre” in Arabic NLP is task- and corpus-dependent rather than a single standardized design category.
7. Limitations, cautions, and research implications
ArzEn-MultiGenre is diverse relative to prior Egyptian Arabic–English parallel datasets, but it remains selective (Al-Sabbagh, 2 Aug 2025). Its songs are all by one artist, Amr Diab; its subtitles come from only two Netflix series; and its novels comprise only three titles (Al-Sabbagh, 2 Aug 2025). The paper therefore cautions against treating the dataset as fully representative of all Egyptian Arabic registers, dialect subvarieties, or social classes (Al-Sabbagh, 2 Aug 2025).
The subtitle subset has an additional sociolect bias. The paper notes that Paranormal and Finding Ola feature well-educated, upper-class Egyptians who often mix English and Egyptian Arabic (Al-Sabbagh, 2 Aug 2025). This implies that the subtitle data may reflect a particular social register and interactional style rather than a comprehensive cross-section of spoken Egyptian Arabic.
Orthographic variation is both a strength and a difficulty. Because spelling and punctuation were intentionally preserved, users should expect spelling variation, punctuation inconsistency, and associated challenges for automated preprocessing and model training (Al-Sabbagh, 2 Aug 2025). This suggests that experiments on tokenization, normalization, subword modeling, and robustness to non-standard script variation are especially relevant when using this corpus.
The paper also identifies a practical issue in downstream MT customization. Google AutoML Translation or Cloud Translation automatically removes segment pairs in which the source segment is identical even if the target differs (Al-Sabbagh, 2 Aug 2025). This is why the number of segment pairs used in the AutoML experiments is lower than the totals reported in the corpus statistics tables (Al-Sabbagh, 2 Aug 2025). For benchmarking reproducibility, this matters because the released corpus size and the effective training size in commercial adaptation workflows need not coincide.
Finally, the source provenance of the materials introduces copyright and reuse considerations. Parts of the corpus derive from commercial songs, published novels, and Netflix shows, and while the dataset is publicly deposited on Mendeley Data under DOI 10.17632/6k97jty9xg.4, the paper excerpt does not spell out a separate dataset-specific license beyond repository availability (Al-Sabbagh, 2 Aug 2025). This suggests that technical reuse is straightforward, whereas redistribution and downstream commercial exploitation may require case-by-case legal scrutiny.
Taken together, these constraints do not diminish the corpus’s importance; rather, they define its research profile. ArzEn-MultiGenre is best understood as a curated, expert-aligned, dialectal Egyptian Arabic–English parallel dataset whose primary scholarly value lies in bringing underrepresented genres into parallel-corpus research while preserving the orthographic and stylistic particularities of real Egyptian Arabic source materials (Al-Sabbagh, 2 Aug 2025).
