sentseg-camelbert-PA

Model Description

sentseg-camelbert-PA is an Arabic sentence segmentation model built by fine-tuning CAMeLBERT-MSA on the Paragraph-Aware (PA) variant of AraSeg, a manually annotated, genre-diverse Arabic sentence segmentation benchmark covering 8 genres.

Sentence segmentation is framed as binary token classification: for every word, the model predicts whether a sentence boundary follows it.

---

Task Variant

Paragraph-Aware (PA): Punctuation and paragraph boundaries are kept; each paragraph is segmented on its own.

AraSeg defines four task variants, each with its own model:

Model	Variant	Paragraph boundaries	Punctuation
sentseg-camelbert-NP	No-Paragraph (NP)	✗	✓
sentseg-camelbert-PA	Paragraph-Aware (PA)	✓	✓
sentseg-camelbert-NoPnx-NP	No-Punctuation No-Paragraph (NoPnx-NP)	✗	✗
sentseg-camelbert-NoPnx-PA	No-Punctuation Paragraph-Aware (NoPnx-PA)	✓	✗

---

Labels

ID	Label	Meaning
0	O	no sentence boundary after this word
1	EOS	a sentence ends after this word

The label of a word is predicted on its last subtoken.

---

Training

• Base model: CAMeL-Lab/bert-base-arabic-camelbert-msa (the tokenizer is unchanged).
• Data: the train split of the PA config of MBZUAI/AraSEG.
• Hyperparameters: effective batch size 32, 5 epochs, learning rate 3e-5.
• Long inputs: sequences are processed in 512-token windows with a stride of 64; logits of overlapping windows are averaged.

---

Evaluation

Boundary precision (P), recall (R) and F1 on AraSeg. A predicted boundary is correct only if it exactly matches a gold boundary; scores are computed per document and macro-averaged over documents.

Split	P	R	F1
Dev	96.3	95.3	95.6
Test	96.1	94.7	95.2

---

Citation

If you use this model 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"
}