BAREC-10M Corpus v1.0

Corpus Summary

BAREC-10M is an expanded version of the Balanced Arabic Readability Evaluation Corpus (BAREC), scaling from 1 million to 10 million words and broadening its scope to include balanced, multi-domain coverage. Each text is labeled by domain, genre, and readership level, and enriched with automatic morphological, syntactic, and readability analysis using state-of-the-art tools.

---

Available Annotations

The corpus includes both document-level and sentence-level annotations.

Document-level annotations (manually labeled):

• Domain: Arts & Humanities, Social Sciences, or STEM
• Readership Group: Foundational, Advanced, or Specialized
• Text Category: Educational Materials, Literature, Art & Music, Media & Culture, Academic, Encyclopedic, or Religion & Philosophy

Sentence-level annotations (automatically generated):

• Morphological analysis
• Syntactic parsing
• Readability leveling

---

Languages

• Arabic (Modern Standard Arabic)

---

Corpus Details

The structure of the dataset directory is as follows:

.
├── Data/
│   ├── Metadata.xlsx
│   ├── Raw.zip
│   ├── Morphology_and_Readability.zip
│   ├── Syntax_CATiB.zip
│   └── Syntax_UD.zip
└── README.md

Metadata

The metadata file contains the following fields:

• Document: Document file name (without extension)
• Directory: Document directory
• Source: Document source
• Book: Book title
• Author: Author name
• Domain
• Readership Level
• Text Category
• Word Count: Number of words in the document
• Sentence Count: Number of sentences in the document
• In BAREC Corpus?: Indicates whether the document originates from the original BAREC corpus (Yes or No)

Raw Sentences

The corpus includes 20,535 .txt files containing raw sentences, organized into multiple directories according to the metadata.

Morphology and Readability

The corpus includes 20,535 .json files containing morphological and readability annotations, organized into multiple directories according to the metadata.

Each JSON file represents a document and contains the following key-value pairs:

Sentence-level features:

• raw_sents: Raw sentences (list of strings)
• sents_word_count: Number of words per sentence (list of integers)
• sents_RL: Sentence-level readability scores (list of integers from 1 to 19). The value ### indicates problematic sentences in documents originating from the BAREC corpus.

Word-level features:

• word: Tokenized words for all sentences (list of lists of strings)
• lex: Lemmas of all words (list of lists of strings)
• pos: Part-of-speech tags (list of lists of strings)
• RL: Readability levels of lemmas (list of lists of integers)
• num, gen, mod, etc.: Additional CAMeL Morph features of all words (list of lists of strings)

Syntax

We provide syntactic annotations in both the Columbia Arabic Treebank (CATiB) and Universal Dependencies (UD) schemes.

The corpus includes 20,535 .conllx files per annotation scheme, each containing syntactic annotations and organized into multiple directories according to the metadata.

We recommend using the Palmyra tool for visualization and analysis of these files.

---

Usage

You can download the files manually using the Hub’s user interface, or use snapshot_download to download all files at once.

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="CAMeL-Lab/BAREC-10M",
    repo_type="dataset",
    local_dir="path/to/local/dir",
    allow_patterns=["Data/*"]
)

---

Citation

If you use BAREC-10M in your work, please cite the following paper:

@inproceedings{elmadani2026large,
  author    = {Elmadani, Khalid N. and Wizani, Adel Mahmoud and Taha-Thomure, Hanada and Habash, Nizar},
  title     = {A Large and Balanced Multi-Domain Arabic Corpus Annotated for Morphology, Syntax, and Readability},
  booktitle = {Proceedings of the International Conference on Language Resources and Evaluation (LREC 2026)},
  year      = {2026},
  address   = {Palma, Mallorca, Spain}
}