Quranic Multilingual Corpus
- Quranic multilingual corpus is a structured ecosystem of aligned Quranic Arabic text, translations, transliteration, and audio to support diverse research applications.
- It integrates resources like QuranMorph and Quran-MD, providing detailed morphological annotation and multimodal alignment with precise indexing.
- The corpus facilitates cross-language retrieval, question answering, and domain-adapted modeling, ensuring interoperability across various Islamic textual datasets.
In current computational scholarship, a Quranic multilingual corpus denotes not a single fixed dataset but an ecosystem of aligned resources that connect the original Quranic Arabic to other linguistic, representational, or modality-specific layers at verse, word, or token granularity. Recent work spans lexicon-linked Arabic morphology, verse-parallel multilingual retrieval collections, Arabic–English question-answering corpora, and multilingual multimodal datasets with aligned recitation audio. Within this landscape, QuranMorph provides a manually lemmatized and part-of-speech tagged Arabic backbone for the entire Qur’an (Akra et al., 22 Jun 2025), Quran-MD supplies Arabic text, English translation, transliteration, and aligned audio at verse and word levels (Salman et al., 25 Jan 2026), multilingual retrieval work treats verse-aligned Arabic, English, Russian, and Urdu collections as a common evaluation substrate (Pavlova et al., 18 Sep 2025), and question-answering datasets attach contextual tafsir and hadith layers or cross-language query formulations to Quranic passages (Oshallah et al., 29 Jan 2025, Qamar et al., 2024).
1. Conceptual scope and representative resource types
The phrase “Quranic multilingual corpus” covers several distinct resource classes. Some corpora are multilingual in the narrow sense of parallel text across languages; others are Arabic-first corpora deliberately designed for interoperability with multilingual lexicons or downstream cross-lingual systems. QuranMorph is explicit on this point: it is a morphologically annotated Arabic corpus of the entire Qur’an, not itself a multilingual corpus, but it is tightly coupled to Qabas and thereby positioned as a backbone for broader multilingual or cross-lingual work (Akra et al., 22 Jun 2025). By contrast, Quran-MD is explicitly described as a fine-grained multilingual, multimodal dataset of the Quran, pairing Arabic with English and phonetic transliteration and aligning both text and audio (Salman et al., 25 Jan 2026).
A compact way to situate the main resource types is to distinguish them by their primary alignment target and intended use.
| Resource | Core content | Role in a multilingual ecosystem |
|---|---|---|
| QuranMorph | Arabic corpus of the entire Qur’an; 77,429 tokens; lemma + 40-tag POS | Lexicon-linked morphological backbone (Akra et al., 22 Jun 2025) |
| Quran-MD | Arabic, English, transliteration, verse- and word-level audio | Multilingual multimodal alignment (Salman et al., 25 Jan 2026) |
| Quran-centered MLIR corpus | Arabic, English, Russian, Urdu verse collections + QRCD queries | Verse-level multilingual retrieval benchmark and deployment substrate (Pavlova et al., 18 Sep 2025) |
| Quranic QA corpora | Arabic–English passage retrieval sets; English long-form QA with tafsir/hadith context | Cross-language QA and contextual reasoning layers (Oshallah et al., 29 Jan 2025, Qamar et al., 2024) |
These resources differ substantially in what they mean by “multilingual.” In Quran-MD, multilinguality currently means Arabic–English text plus transliteration, with an architecture designed for future language expansion (Salman et al., 25 Jan 2026). In multilingual retrieval, the corpus is a verse-parallel four-language collection used to test monolingual, cross-lingual, and multilingual search over a common set of verse IDs (Pavlova et al., 18 Sep 2025). In morphologically centered work, multilinguality often arises indirectly through shared lemma inventories, multilingual lexicographic links, or alignment to translated corpora rather than through translations embedded in the corpus itself (Akra et al., 22 Jun 2025).
2. Alignment units, indices, and corpus data models
Verse-level alignment is the dominant interoperability mechanism across Quranic multilingual resources. In the multilingual retrieval setting, each verse is treated as a document or passage, and relevance judgments from the Arabic Qur’anic Reading Comprehension Dataset are mapped to the corresponding verses in English, Urdu, and Russian because the corpus is parallel at the verse level (Pavlova et al., 18 Sep 2025). The cross-language Quranic QA work follows the same principle: Arabic questions are translated into English, relevant answers are retrieved from an English translation of the Quran, and verse alignment is preserved through surah and ayah indexing (Oshallah et al., 29 Jan 2025).
Word-level alignment becomes central when a corpus aims to support lexical, phonological, or multimodal analysis. QuranMorph adopts the segmentation logic of the earlier Quranic Arabic Corpus but aggregates clitics into full words and indexes each token with a location of the form (chapter:verse:word), such as (1:1:1). This allows seamless integration with the Quranic Arabic Corpus, whose finer unit is (chapter:verse:word:clitic) (Akra et al., 22 Jun 2025). Quran-MD implements a hierarchical JSON structure with three nested levels—surah, ayah, and word—and preserves deterministic ordering across Arabic text, English translation, transliteration, and audio paths. At the word level, each token is paired with word_arabic, word_english, word_transliteration, and audio_word_path (Salman et al., 25 Jan 2026).
For morphologically annotated Arabic tokens, the representation can be summarized as
where is a Qabas lemma, belongs to the SAMA/Qabas inventory, and is the chapter/verse/word index (Akra et al., 22 Jun 2025). For multilingual retrieval, the comparable abstraction is a verse embedding space in which a query and verse are scored by cosine similarity,
with verse IDs serving as the alignment anchor across languages (Pavlova et al., 18 Sep 2025).
This architecture has an important consequence: Quranic multilingual corpora are usually compositional rather than monolithic. A plausible implication is that interoperability depends less on a single universal format than on stable keys such as surah, ayah, and word position, because these keys permit the later fusion of morphology, translation, transliteration, tafsir, hadith, and audio layers drawn from different projects.
3. Morphological annotation as the Arabic backbone
QuranMorph is the most explicit attempt to provide a standardized Arabic linguistic substrate for multilingual or cross-resource use. It covers the entire Qur’an with 77,429 words (tokens) in 6,235 verses, including 19,009 unique surface words and 4,616 unique lemmas. The lemma inventory is further divided into 3,057 noun lemmas, 1,479 verb lemmas, and 173 functional word lemmas. Each token is manually lemmatized and tagged for part of speech by three expert linguists using lemmas from Qabas and the fine-grained SAMA/Qabas POS inventory of 40 tags (Akra et al., 22 Jun 2025).
The annotation workflow is manual but tool-assisted. The corpus is loaded into the Tawseem web annotation tool, which integrates the ALMA morphological tagger to suggest lemmas and POS tags; annotators then review each suggestion in full verse context. If no suitable lemma exists in Qabas, a new lemma entry is added with linguistic and morphological details, so the QuranMorph annotation layer and the shared Qabas lexicographic database remain tightly coupled (Akra et al., 22 Jun 2025). Qabas itself is described as containing about 60,000 lemmas, linked to 110 lexicons and about 2 million tokens across Classical Arabic, Modern Standard Arabic, and dialect corpora, which makes QuranMorph lemmas immediately reusable in broader lexicographic and corpus contexts (Akra et al., 22 Jun 2025).
A key design decision is compatibility with mainstream Arabic NLP infrastructure. The earlier Quranic Arabic Corpus used a specialized POS tagset of 44 tags tailored to Quranic and traditional grammar, and its segmentation operated at clitic level. QuranMorph instead uses the same 40 POS tags as SAMA/Qabas and retains compatibility with the Quranic Arabic Corpus through shared tokenization logic and location indices (Akra et al., 22 Jun 2025). Because corpora such as PADT, PATB, Curras, Baladi, Nabra, Gumar, and Lisan are also annotated using SAMA or Qabas lemmas and tagsets, QuranMorph can be placed in a unified lemma space that spans Classical, Modern Standard, and dialectal Arabic (Akra et al., 22 Jun 2025).
The corpus is also explicitly context-sensitive in its lemmatization policy. Annotators are instructed to read the full verse and, where necessary, consult tafsīr, because Quranic usage may diverge from Modern Standard Arabic senses. The paper gives examples such as عَفَا in Qur’an 7:95, annotated with the lemma meaning “grew/increased” rather than the more common Modern Standard Arabic sense “forgave/pardoned,” and ضَرَب, whose contextual meaning varies across verses (Akra et al., 22 Jun 2025). For multilingual corpus design, this matters because translation alignment anchored in contextually accurate lemmas is semantically stronger than alignment based only on surface forms.
4. Multilingual and multimodal realization
Quran-MD is the clearest example of a Quranic corpus that is multilingual by construction and multimodal by design. It covers the complete Quran with 114 surahs, 6,236 ayahs, and approximately 77.8k words. At verse level, each ayah includes fully diacritized Arabic text, an English translation, a phonetic transliteration, and verse-level audio. At word level, each token is represented by Arabic script, an English word-level translation, transliteration, and an aligned audio clip (Salman et al., 25 Jan 2026).
The dataset’s audio layer is unusually dense. Verse-level recordings come from the Kaggle Quran Ayat Speech-to-Text Dataset, with the main dataset statistics and description reporting 30 reciters and about 665 hours of verse-level audio, while the abstract states 32 distinct reciters. Word-level audio comes from the Internet Archive Quran Word-by-Word Audio Collection and totals about 22 hours (Salman et al., 25 Jan 2026). The paper does not resolve the 30-versus-32 discrepancy; this suggests a difference between the released dataset statistics and the abstract’s summary count.
The multilingual layer is currently bilingual rather than broadly multilingual. Quran-MD contains Arabic and English, plus transliteration derived from quranwbw resources. It does not include explicit POS tags, lemmas, morphological features, or syntactic dependencies, and it does not yet provide timestamps within verse audio, phoneme-level segmentation, or tajweed labels such as ghunnah, madd, or idgham (Salman et al., 25 Jan 2026). Alignment is deterministic and index-based rather than learned: word order is preserved across Arabic, English, transliteration, and audio, and validation scripts are used to detect missing or misaligned entries (Salman et al., 25 Jan 2026).
This resource demonstrates that “multilingual” in Quranic corpus work increasingly intersects with “multimodal.” Quran-MD positions itself against two adjacent traditions: morpho-syntactic corpora such as the Quranic Arabic Corpus and MASAQ, which provide deeper linguistic annotation but limited or no audio, and translation-heavy resources such as Tanzil, which offer >40-language translations but no audio or word-level multimodal alignment (Salman et al., 25 Jan 2026). The result is a structurally different kind of corpus, one aimed at text–speech integration, semantic retrieval, recitation analysis, and educational technology rather than morphological analysis alone.
5. Retrieval, question answering, and domain-adapted modeling
Verse-aligned multilingual corpora have become a practical substrate for retrieval and question answering. In multilingual information retrieval, a Quran-centered collection in Arabic, English, Russian, and Urdu is used alongside translated QRCD queries to evaluate monolingual IR, cross-lingual IR, and multilingual IR within one framework. The corpus is verse-parallel; each verse in Arabic has a corresponding verse in each translation, and relevance labels are mapped by verse ID across languages. On 169 test queries, the mixed-training model Biq-ENc achieved the best average monolingual MRR@10 of 0.426, while 4lingual-train-all achieved the best average multilingual MRR@10 of 0.357, narrowly ahead of Biq-ENc at 0.354 (Pavlova et al., 18 Sep 2025). The same work also reports deployment of a single lightweight model as a multilingual Quran search tool, emphasizing latency and cost-efficiency in real-world MLIR (Pavlova et al., 18 Sep 2025).
Cross-language Quranic QA uses the Quran QA 2023 passage retrieval setting as a seed and expands it into an Arabic–English corpus. The authors start from 251 Arabic questions, add external Quran-related question sources, and obtain 629 distinct Arabic questions. Each question is paraphrased twice, producing a final set of 1,895 questions, including single-answer, multi-answer, and zero-answer cases. These questions are translated from Arabic into English using the Google Translate API, and retrieval is then performed over Pickthall’s English Quran translation, with verse alignment preserved by surah and ayah indexing. In this setting, RoBERTa-Base achieved the highest MAP@10 (0.34) and MRR (0.52), while DeBERTa-v3-Base achieved the best Recall@10 (0.50) and Precision@10 (0.24) (Oshallah et al., 29 Jan 2025).
Long-form Quranic QA introduces a different corpus form: not verse-parallel translation, but contextualized expert question–answer pairs connected to Quranic tafsir and hadith. One benchmark is built from Islamqa.org, an English translation of Tanwīr al-Miqbās min Tafsīr Ibn ‘Abbās, and English translations of more than 33,000 Ahadith from the six canonical collections. The abstract reports over 73,000 question-answer pairs, while the topic evaluation table lists 77,113 questions. Each row contains a Question, Answer, and Context, where context is formed by selecting the top 3 tafsir ayahs and top 3 Ahadith from the same topic using soft cosine similarity (Qamar et al., 2024). This dataset is effectively English in its operational form, because Arabic answer counterparts are removed during preprocessing (Qamar et al., 2024).
Evaluation methodology is a central issue in this long-form setting. Automatic scores improved after fine-tuning—for example, LED reached ROUGE-1 = 27.23, ROUGE-2 = 7.55, ROUGE-L = 18.13, and BERTScore = 78.95—yet human evaluation showed that verdict consistency with expert scholars ranged only from 11% to 22%, while contextual understanding ranged from 53% to 90% across models (Qamar et al., 2024). This indicates that a multilingual Quranic corpus used for QA cannot be evaluated adequately by lexical overlap alone when doctrinal correctness is part of the task.
Domain adaptation further extends the notion of a Quranic multilingual corpus beyond verse-aligned translations. In Qur’anic IR for Arabic and English, an English Islamic corpus of about 47M words is compiled from Islamic literature, journals, fatwa counseling, Islam-related Wikipedia articles, and filtered Common Crawl, and is used to continue pre-train a domain-specific English LLM, BPIT. The adapted vocabulary adds 3992 new domain-specific tokens and yields 33,511 total tokens (Pavlova, 2023). Combined with Tafseer Ibn Kathir–based augmentation, this improves verse-level retrieval: in English, SBERT-ID reaches MRR@10 = 0.55, NDCG@5 = 0.33, and Recall@100 = 0.32; in Arabic, ColBERT-AR-ID reaches MRR@10 = 0.48, NDCG@5 = 0.29, and Recall@100 = 0.29 (Pavlova, 2023). The broader implication is that multilingual Quranic corpus construction increasingly includes domain-adapted background corpora, not only verse translations.
6. Limitations, misconceptions, and future directions
A persistent misconception is that any Quranic multilingual corpus must already contain many natural languages. Current resources show a more fragmented reality. Quran-MD is currently Arabic + English plus transliteration, despite being designed for future language expansion (Salman et al., 25 Jan 2026). The cross-language QA corpus is effectively Arabic–English and relies on machine translation of questions into English (Oshallah et al., 29 Jan 2025). The deployed MLIR study uses four languages—Arabic, English, Russian, and Urdu—even though the paper notes that the Quran has translations in over 100 languages (Pavlova et al., 18 Sep 2025). Multilinguality, in practice, is often constrained by availability of verified translations, evaluation sets, and compatible indexing.
A second misconception is that multilinguality and linguistic depth normally co-occur. The opposite pattern is common. QuranMorph offers dense Arabic lemmatization and POS annotation but no embedded translations or multilingual annotations (Akra et al., 22 Jun 2025). Quran-MD offers multilingual text, transliteration, and audio but no explicit lemmas, morphology, POS tags, or syntax (Salman et al., 25 Jan 2026). The long-form QA benchmark offers rich contextual reasoning over tafsir and hadith, but it is operationally monolingual English and lacks explicit verse- or hadith-ID schema in the released task description (Qamar et al., 2024). This suggests that Quranic multilingual corpora are best understood as layered combinations of complementary resources rather than as single all-purpose datasets.
Evaluation remains particularly difficult in religious and interpretive tasks. In retrieval, verse-level alignment and exhaustive relevance judgments make metrics such as MRR@10, nDCG@10, and Success@5 operationally useful (Pavlova et al., 18 Sep 2025). In long-form QA, however, the discrepancy between automatic metrics and expert judgment is stark: high ROUGE overlap does not guarantee correct scholarly verdicts (Qamar et al., 2024). For cross-language Quranic QA, machine translation and paraphrasing enlarge the dataset but also introduce fidelity risks that the paper does not systematically validate by human review (Oshallah et al., 29 Jan 2025). In sacred-text applications, this is not a marginal issue.
The most consistent future directions are additive rather than substitutive. Quran-MD explicitly proposes more languages, forced alignment, phoneme-level segmentation, timestamps, and tajweed annotation (Salman et al., 25 Jan 2026). QuranMorph is already designed to integrate with the Quranic Arabic Corpus, Qabas, and other SAMA/Qabas-based corpora, making further multilingual lexicon linkage a natural extension (Akra et al., 22 Jun 2025). Multilingual retrieval work points toward broader Islamic-text MLIR over tafsir and hadith beyond the Quran alone (Pavlova et al., 18 Sep 2025). Long-form QA work suggests attaching Arabic texts, TEI-encoded resources, and hadith knowledge-graph structures to English contextual QA datasets (Qamar et al., 2024). A plausible implication is that the mature form of a Quranic multilingual corpus will be federated: verse-aligned across languages, word-aligned where possible, lexicon-linked on the Arabic side, and incrementally enriched with recitation audio, tafsir, hadith, and evaluation protocols that can accommodate both semantic relevance and doctrinal correctness.