Arabic Sentence Segmentation Shared Task 2026

For details about the shared task, evaluation scripts, leaderboard, and submission guidelines, visit: https://www.araseg.aramlab.ai/

Dataset Summary

AraSeg is the first comprehensive benchmark for Arabic sentence segmentation. The corpus is designed to support research on sentence segmentation in Modern Standard Arabic (MSA), particularly in settings where punctuation is inconsistent, missing, or noisy. AraSeg contains manually annotated documents collected from diverse sources and genres, enabling robust evaluation across different writing styles and domains.

AraSeg-NP is the No-Paragraph (NP) variant of the corpus where documents include punctuation marks but no paragraph boundaries.

---

Dataset Structure

Data Instances

{'doc_id': 'doc_00b450a96684',
'tokens': ['الفصل','الأول','حين','ركبت','السيارة','لم','أكن','أتصور','أنني','أبدأ', ...],
'labels': [0, 1, 0, 0, 0, 0, 0, 0, 0, 0, ...],
'text': 'أبدأ الفصل الأول حين ركبت السيارة لم أكن أتصور أنني...',
'label_str': '0100000000...'}

Data Fields

• doc_id: Unique document identifier.
• tokens: White-space-tokenized document represented as a list of tokens.
• labels: Token-level sentence boundary labels. 1 indicates that a sentence boundary follows the current token, while 0 indicates no boundary.
• text: Document.
• label_str: Sentence boundary labels as a binary string. 1 indicates that a sentence boundary follows the current token, while 0 indicates no boundary.

Data Splits

• train: 174 documents (10,657 sentences and 124K words).
• dev: 222 documents (12,985 sentences and 159K words).
• test: 262 documents (12,509 sentences and 154K words).

---

Task Definition

Sentence segmentation is formulated as a binary token classification task.

Given a sequence of tokens: [token_1, token_2, ..., token_n], the model predicts for each token whether a sentence boundary follows it. For example:

The label 1 indicates that the sentence ends after the token.

---

Evaluation

We evaluate systems using boundary-level metrics:

• Boundary Precision (P): Percentage of predicted sentence boundaries that are correct.
• Boundary Recall (R): Percentage of gold sentence boundaries correctly identified.
• Boundary F1 (F1): Harmonic mean of precision and recall.

Metrics are computed at the document level and averaged across the corpus.

We provide evaluation scripts on this repo.

---