'ARIQ is the first manually annotated dataset dedicated to the extraction and attribution of Islamic quotations in Arabic legal discourse (Fatwas). It provides a fine-grained, frame-based annotation schema covering 14 distinct tags across three primary quotation source types: the Qur'an, Hadith, and scholarly citations.

Dataset Overview

'ARIQ addresses domain-specific challenges in religious and legal text processing, including complex nested quotations (up to 5 levels deep), multi-part question-and-answer exchange structures, and multiple attribution components.

Total Documents: 1,833 fatwa-style answers (1,473 training / 360 test)
Total Paragraphs: 4,980
Total Words: ~855,170 words
Total Annotated Quotation Instances: 8,789
Annotation Agreement: Near-perfect inter-annotator agreement with a global Cohen's Kappa of 0.935.
Data Sources: IslamQA (1,153 documents) and Islamweb (680 documents).

Dataset Structure & File Formats

'ARIQ is released in two complementary formats to support diverse Natural Language Processing (NLP) pipelines: Span-Based Formats (JSON) and Sequence Labeling Formats (Parquet). Both train and test splits are provided for each format.

1. Span-Based Format (JSON)

Represented as labeled spans with start/end character offsets and annotation types. Provided in two versions:

Document-Level (train_doc.jsonl / test_doc.jsonl): Preserves the complete fatwa text per record along with global span labels.
Paragraph-Level (train_para.jsonl / test_para.jsonl): Organizes annotations granularly by paragraph.

Example JSON Structure: for Document-Level

{
  "DocID": 2307,
  "Text": "... فقال : ( لا يُؤَاخِذُكُمُ الله ...",
  "Label": [
    [143, 500, "quran_quotation_frame"],
    [151, 487, "quran_content"],
    [488, 495, "quran_chapter_name"],
    [498, 500, "quran_verse_number"]
  ],
  "Src": "IslamQA",
  "SrcQuestion": "كانت عليَّ كفارة يمين ، فصمت ثلاثة أيام مع قدرتي على إطعام عشرة مساكين ، فهل ما فعلته صحيح ؟.",
  "SrcCtg": [
    ["الفقه وأصوله", "الفقه", "عبادات", "الأيمان والنذور"]
  ],
  "UnifCtg": [
    ["الأيمان والنذور", "الأيمان", "أحكام اليمين والنذر"]
  ],
  "SrcUrl": "https://islamqa.info/ar/answers/42804/"
}

Example JSON Structure: for Paragraph-Level

{
  "DocID": 2307,
  "Paras": [
    {
        "ParaID": 3661,
        "ParaText": "... فقال : ( لا يُؤَاخِذُكُمُ اللَّهُ ...",
        "ParaLabel": [
            [143, 500, "quran_quotation_frame"],
            [151, 487, "quran_content"],
            [488, 495, "quran_chapter_name"],
            [498, 500, "quran_verse_number"]
        ]
    }
  ],
  "Src": "IslamQA",
  "SrcQuestion": "كانت عليَّ كفارة يمين ، فصمت ثلاثة أيام مع قدرتي على إطعام عشرة مساكين ، فهل ما فعلته صحيح ؟.",
  "SrcCtg": [
    ["الفقه وأصوله", "الفقه", "عبادات", "الأيمان والنذور"]
  ],
  "UnifCtg": [
    ["الأيمان والنذور", "الأيمان", "أحكام اليمين والنذر"]
  ],
  "SrcUrl": "https://islamqa.info/ar/answers/42804/"
}

2. Sequence-Labeling Format (Parquet)

Represented as tokenized word sequences with IOB-based tags assigned across multiple hierarchical levels to handle nesting and overlapping components. Provided as train_seq.parquet and test_seq.parquet.

Example CSV Structure:

DocID	ParaID	Word	Tag-1	Tag-2
2307	3661	فقال	B-quran_quotation_frame
2307	3661	(	I-quran_quotation_frame	B-quran_content
2307	3661	لا	I-quran_quotation_frame	I-quran_content
2307	3661	يُؤَاخِذْكُمُ	I-quran_quotation_frame	I-quran_content
2307	3661	المائدة	I-quran_quotation_frame	B-quran_chapter_name
2307	3661	89	I-quran_quotation_frame	B-quran_verse_number

Loading the dataset

This dataset is gated. Request access, then log in with huggingface-cli login.

from datasets import load_dataset

ds_doc  = load_dataset("Misraj/ARIQ", "span_doc")    # document-level spans (JSONL)
ds_para = load_dataset("Misraj/ARIQ", "span_para")   # paragraph-level spans (JSONL)

# Sequence labeling (Parquet): load with the parquet builder explicitly
base = "hf://datasets/Misraj/ARIQ/seq"
ds_seq = load_dataset("parquet", data_files={
    "train": f"{base}/train_seq.parquet",
    "test":  f"{base}/test_seq.parquet",
})

Annotation Taxonomy (14 Tags)

The annotation schema comprises three quotation-frame types, three content types, and eight source-attribution types:

Qur'anic Quotations:
- quran_quotation_frame
- quran_content
- quran_chapter_name (Surah Name)
- quran_verse_number (Ayah Number)
Hadith Quotations:
- hadith_quotation_frame
- hadith_content
- hadith_narrator_name (Narrator / Speaker)
- hadith_referenced_book
- hadith_number
Scholarly Quotations:
- scholar_quotation_frame
- scholar_content
- scholar_name (Scholar / Jurist)
- scholar_referenced_book
- scholar_page_number

Data Split Statistics

The dataset uses a stratified 80/20 split preserving tag distributions and platform sources across partitions:

Training Set: 1,473 documents
Test Set: 360 documents

Baseline Performance

Baseline evaluation using fine-tuned AraBERTv0.2-base models under category-specific and joint unified paradigms achieved strong component-level macro-$F_1$ scores (reaching 0.8583 overall).

Potential Use Cases

While the corpus is primarily designed and structured for quotation extraction, span-based identification, and hierarchical sequence labeling, its rich metadata and textual contents can support auxiliary downstream tasks, including:

Attribution Source Identification: Training models to identify and link specific speakers, narrators, or legal authorities (such as scholars, books, or scriptural chapters) to their corresponding statements.
Extractive Question Answering (QA): Leveraging the natural question-and-answer structure of fatwa documents—along with their precise scriptural quotations—for exploratory QA or retrieval-augmented generation (RAG) tasks in the religious domain.
Legal and Religious Document Classification: Leveraging the fatwa text and platform source metadata for text categorization or domain-specific classification tasks within Islamic jurisprudence.
Hierarchical Boundary Precision: Benchmarking models on complex, nested text spans to evaluate how well architectures handle multi-level token dependencies and overlapping annotations.

Citation

If you use 'ARIQ in your research, please cite our paper:

@inproceedings{kheelan2026ariq,
  title={{'ARIQ}: A Dataset for Quotation Extraction and Attribution in Islamic Jurisprudential Texts},
  author={Kheelan, Islam and others},
  organization={Misraj AI},
  booktitle={Proceedings of the 9th International Conference on Arabic Language Processing (ICALP)},
  year={2026},
  publisher={Springer}
}

Downloads last month: -

Total file size:

39.5 MB