The Arabic Sentence Segmentation Shared Task 2026 will take place at the Fourth Arabic Natural Language Processing Conference (ArabicNLP 2026) at EMNLP 2026.
Click here to register for the shared task!
The goal of this shared task is to segment Arabic documents into sentences. Given a document, systems must predict sentence boundaries to produce coherent sentences. The task is formulated as a binary classification task, where models predict whether a boundary follows each token.
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. The dataset has four variants that differ in terms of paragraph structure and punctuation availability:
The shared task features four subtasks, each corresponding to an AraSeg corpus variant, with two tracks. The tracks impose different resource constraints. Participants can compete in one or more subtask and track.
Closed Track: Models must be trained exclusively on the AraSeg corpus.
Open Track: No restrictions on external resources, allowing the use of any publicly available data.
With four subtasks and two tracks, the task results in eight possible combinations. Participants are allowed to compete in multiple subtasks and tracks.
We define sentence segmentation as a binary token classification task. We use the following metrics:
We compute these metrics per document, and average across all documents. We provide instructions on how to run the evaluation script below.
You will need to have conda installed. To setup the environment, you would need to run:
git clone https://github.com/mbzuai-nlp/araseg-shared-task-2026
cd araseg-shared-task-2026
conda create -n araseg python=3.10
conda activate araseg
pip install -r requirements.txt
To evaluate your predictions, use the provided evaluation script. The script requires three arguments:
--predictions: Path to your output CSV file containing predictions--task: The data split to evaluate on (dev or test).--split: The segmentation task type (PA for Paragraph-Aware, NP for No-Paragraph, NoPnx-PA for No-Punctuation Paragraph-Aware, and NoPnx-NP for No-Punctuation No-Paragraph).To evaluate your system's output, you would need to run:
python scripts/eval.py --predictions /path/to/prediction_csv --task [PA|NP|NoPNX-PA|NoPNX-NP] --split [dev|test]
Example usage:
python scripts/eval.py --predictions examples/NoPnx-NP_dev.csv --task NoPnx-NP --split dev
Your output CSV file should have the following columns:
Document ID: The unique identifier for each document.Prediction: Your predicted segmentation boundaries for each as token in the document. This must be represented as a binary sequence of 0s and 1s (e.g., 0010 for a document of 4 tokens).Example:
| Document ID | Prediction |
|---|---|
| doc_00f88da2b078 | 00111001001000001011 |
| doc_4f96c619afba | 011010111010101 |
| ... | ... |
Make sure the IDs in your output file match exactly those in the provided split (dev or test) for the chosen task.
After running the evaluation script, you will see output similar to the following in your terminal:
--------------------
Prediction file path: examples/NoPnx-NP_dev.csv
Task name: NoPnx-NP
Split: dev
--------------------
Precision: 43.19
Recall: 88.19
F1: 52.54
--------------------
Each metric reflects the performance of your predictions on the selected split and task.
This repo is available under the MIT license. See the LICENSE for more info.
Arabic Sentence Segmentation Across Genres and Punctuation Conditions. Mohammed Elkholy, Khalid N. Elmadani, Nizar Habash, and Bashar Alhafni. arXiv Preprint.
The Arabic Sentence Segmentation Shared Task 2026 will take place at the Fourth Arabic Natural Language Processing Conference (ArabicNLP 2026) at EMNLP 2026.
Click here to register for the shared task!
The goal of this shared task is to segment Arabic documents into sentences. Given a document, systems must predict sentence boundaries to produce coherent sentences. The task is formulated as a binary classification task, where models predict whether a boundary follows each token.
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. The dataset has four variants that differ in terms of paragraph structure and punctuation availability:
The shared task features four subtasks, each corresponding to an AraSeg corpus variant, with two tracks. The tracks impose different resource constraints. Participants can compete in one or more subtask and track.
Closed Track: Models must be trained exclusively on the AraSeg corpus.
Open Track: No restrictions on external resources, allowing the use of any publicly available data.
With four subtasks and two tracks, the task results in eight possible combinations. Participants are allowed to compete in multiple subtasks and tracks.
We define sentence segmentation as a binary token classification task. We use the following metrics:
We compute these metrics per document, and average across all documents. We provide instructions on how to run the evaluation script below.
You will need to have conda installed. To setup the environment, you would need to run:
git clone https://github.com/mbzuai-nlp/araseg-shared-task-2026
cd araseg-shared-task-2026
conda create -n araseg python=3.10
conda activate araseg
pip install -r requirements.txt
To evaluate your predictions, use the provided evaluation script. The script requires three arguments:
--predictions: Path to your output CSV file containing predictions--task: The data split to evaluate on (dev or test).--split: The segmentation task type (PA for Paragraph-Aware, NP for No-Paragraph, NoPnx-PA for No-Punctuation Paragraph-Aware, and NoPnx-NP for No-Punctuation No-Paragraph).To evaluate your system's output, you would need to run:
python scripts/eval.py --predictions /path/to/prediction_csv --task [PA|NP|NoPNX-PA|NoPNX-NP] --split [dev|test]
Example usage:
python scripts/eval.py --predictions examples/NoPnx-NP_dev.csv --task NoPnx-NP --split dev
Your output CSV file should have the following columns:
Document ID: The unique identifier for each document.Prediction: Your predicted segmentation boundaries for each as token in the document. This must be represented as a binary sequence of 0s and 1s (e.g., 0010 for a document of 4 tokens).Example:
| Document ID | Prediction |
|---|---|
| doc_00f88da2b078 | 00111001001000001011 |
| doc_4f96c619afba | 011010111010101 |
| ... | ... |
Make sure the IDs in your output file match exactly those in the provided split (dev or test) for the chosen task.
After running the evaluation script, you will see output similar to the following in your terminal:
--------------------
Prediction file path: examples/NoPnx-NP_dev.csv
Task name: NoPnx-NP
Split: dev
--------------------
Precision: 43.19
Recall: 88.19
F1: 52.54
--------------------
Each metric reflects the performance of your predictions on the selected split and task.
This repo is available under the MIT license. See the LICENSE for more info.
Arabic Sentence Segmentation Across Genres and Punctuation Conditions. Mohammed Elkholy, Khalid N. Elmadani, Nizar Habash, and Bashar Alhafni. arXiv Preprint.