AraSeg

Dataset Summary

AraSeg is a manually annotated, genre-diverse benchmark for Arabic sentence segmentation in Modern Standard Arabic. It contains 2,187 documents with 48,250 paragraphs and 121,623 sentences (about 1.45M tokens), drawn from 30 sources spanning 8 genres: Children, Education, Literature, Media, Poetry, Politics, Religion and Wiki.

Arabic punctuation is sparse, inconsistent and ambiguous, and many sentence boundaries are not marked by punctuation at all: 39% of AraSeg sentence boundaries are word-based. AraSeg covers punctuated and unpunctuated text, with and without paragraph structure, so segmentation can be studied under realistic conditions. Most of the gold segmentations come from the CAMeL Treebank and the BAREC corpus; the full Quran and additional Old and New Testament text were added to broaden the religious genre.

---

Supported Tasks

Sentence segmentation is framed as binary token classification: for every token, predict whether a sentence boundary follows it. The dataset has four configs (task variants) that differ in the availability of punctuation and paragraph boundaries:

Config	Paragraph boundaries	Punctuation
PA (default)	✓	✓
NP	✗	✓
NoPnx-PA	✓	✗
NoPnx-NP	✗	✗

• PA (Paragraph-Aware): paragraph boundaries are kept, as "\n" tokens between paragraphs.
• NP (No-Paragraph): each document is one stream of tokens.
• NoPnx: all punctuation and symbol characters are removed, along with source-specific boundary markers (Quran, Old and New Testament verse numbers; ArabicMMLU answer labels such as ( أ ).

from datasets import load_dataset

ds = load_dataset("MBZUAI/AraSEG", "NoPnx-NP")

---

Languages

• Arabic (Modern Standard Arabic)

---

Dataset Structure

Data Instances

One row per document (lists truncated):

{'doc_id': 'doc_4a2fa5a558fa', 'doc_name': 'ArabicMMLU_Group_10_001', 'source': 'ArabicMMLU', 'genre': 'Education', 'tokens': ['معنى', 'كلمة', 'صاغ', 'في', 'قول', 'الشاعر', 'صاغ', 'نصر', 'العرب', 'في', 'يوم', 'الكرامة', ...], 'labels': [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, ...], 'paragraph_ids': [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...], 'sent_ids': [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...], 'text': 'معنى كلمة صاغ في قول الشاعر صاغ نصر العرب في يوم الكرامة ( أ ) صاغة ( ب ) سطع ( ج ) فزء ( د ) صنع \n معنى كلمة ضرب ...', 'label_str': '0000000000010001000100010001...'}

Data Fields

• doc_id: Stable document identifier (the same ids as the shared-task datasets).
• doc_name: Original document name.
• source: Source corpus of the document (e.g. Majed, Hindawi, Quran, ArabicMMLU).
• genre: One of Children, Education, Literature, Media, Poetry, Politics, Religion, Wiki.
• tokens: Whitespace tokens of the document (punctuation is already separated from words). In the PA configs, a "\n" token marks each paragraph boundary.
• labels: 1 if a sentence ends after the token, else 0. "\n" tokens are always 0.
• paragraph_ids: Paragraph id of each token. A "\n" token carries the id of the paragraph it closes.
• sent_ids: Sentence id within its paragraph for each token (-1 for "\n" tokens).
• text: The tokens joined by single spaces.
• label_str: labels as a string of 0s and 1s, the format used for shared-task submissions.

Data Splits

• AraSeg has three splits: Train (≈80%), Dev (≈10%) and Test (≈10%), by word count.
• The splits are at the document level, to prevent leakage.
• The original CAMeL Treebank and BAREC splits are preserved. The added Quran, Old Testament and New Testament documents were assigned to splits greedily, balancing word counts.
• All four configs contain the same documents in each split.

Genre Statistics

Genre	#Docs	#Paras	#Sents	#Tokens	Pnx Density	Avg. Sent. Len.
Children	422	9.8K	16.7K	172.7K	20.6%	10.4
Education	605	18.8K	50.3K	373.3K	25.3%	7.4
Literature	406	8.0K	21.0K	339.4K	14.2%	16.2
Media	101	1.2K	2.5K	32.9K	12.4%	13.2
Poetry	30	1.7K	1.7K	11.7K	7.7%	6.9
Politics	17	5.9K	9.7K	147.0K	10.6%	15.1
Religion	438	1.4K	13.7K	216.0K	16.1%	15.7
Wiki	168	1.6K	6.1K	128.6K	13.8%	21.2

---

Annotation

AraSeg re-uses the gold sentence segmentation of the CAMeL Treebank and BAREC, which follow the CAMeL Treebank guidelines: automatic splitting on punctuation (, . ; ? ! :), followed by manual correction through controlled splits and merges. Dependent or incomplete clauses are never segmented into separate sentences. For the Quran, the Old and New Testament, and poetry, verse boundaries are sentence boundaries; in poetry they are also paragraph boundaries. For sources without explicit paragraphs, paragraphs follow the document structure (e.g. Quran chapters, ArabicMMLU questions). In ArabicMMLU, the question and each answer option are separate sentences.

---

Evaluation

We define sentence segmentation as a binary token classification task. The following metrics are used for evaluation:

• Boundary Precision (P): The percentage of predicted sentence boundaries that match a reference boundary. Higher precision means fewer false boundary insertions.
• Boundary Recall (R): The percentage of reference sentence boundaries that are predicted. Higher recall means fewer missed boundaries.
• Boundary F1 (F1): The harmonic mean of boundary precision and recall.

A prediction is correct only if it exactly matches a gold boundary. Scores are computed per document and macro-averaged over documents, so long documents do not dominate. The evaluation script is available at https://github.com/mbzuai-nlp/araseg-shared-task-2026.

---

Citation

If you use AraSeg in your work, please cite the following paper:

@inproceedings{elkholy-etal-2026-arabic-sentence-segmentation,
    title = "{A}rabic Sentence Segmentation Across Genres and Punctuation Conditions",
    author = "Elkholy, Mohammed and
      Elmadani, Khalid N. and
      Habash, Nizar and
      Alhafni, Bashar",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2026",
    month = oct,
    year = "2026",
    address = "Budapest, Hungary",
    publisher = "Association for Computational Linguistics"
}