Abjad AI at AraSeg Shared Task 2026: Auxiliary Loss for Arabic Sentence Segmentation

Abjad AI's submission to the AraSeg 2026 Shared Task on Arabic sentence segmentation achieved F1 scores of 91.7 and 89.7 in the blind tests of the Paragraph-Aware (PA) and No-Paragraph (NP) sub-tasks, respectively, according to the team's paper published in the Proceedings of The Fourth Arabic Natural Language Processing Conference (pages 1004–1009, Budapest, Hungary). Rather than substituting a larger encoder, the team retained the CAMeLBERT-MSA reference system and combined boundary labels aligned to each word's final sub-token, an auxiliary classification head on the penultimate encoder layer trained jointly with the final-layer head, class-weighted cross-entropy for boundary sparsity, and overlap-averaged decoding that scores documents of arbitrary length without truncation. The authors, led by Mohammed Alsudairi, ablated each component in isolation, including the two the reference system already uses, to measure remaining headroom in the task's formulation.

Abstract This paper describes the Abjad AI team’s sub- mission to the AraSeg 2026 Shared Task on Arabic sentence segmentation, covering the two punctuated tracks: Paragraph-Aware PA and No-Paragraph NP . Rather than substituting a larger encoder, we retain the CAMeLBERT- MSA reference system and ask how much head- room remains in the task’s formulation. Our system combines four components: boundary labels aligned to each word’s final sub-token, an auxiliary classification head on the penul- timate encoder layer trained jointly with the final-layer head, class-weighted cross-entropy for boundary sparsity, and overlap-averaged decoding that scores documents of arbitrary length without truncation. We ablate each com- ponent in isolation, including the two that the reference system already uses. Our solution achieves F1 scores of 91.7 and 89.7 in the blind tests of the PA and NP sub-tasks, respectively. - Anthology ID: - 2026.arabicnlp-sharedtasks.123 - Volume: - Proceedings of The Fourth Arabic Natural Language Processing Conference: Shared Tasks https://aclanthology.org/volumes/2026.arabicnlp-sharedtasks/ - Month: - October - Year: - 2026 - Address: - Budapest, Hungary - Editors: - Muhammad Abdul-Mageed https://aclanthology.org/people/muhammad-abdul-mageed/ , Tamer Elsayed https://aclanthology.org/people/tamer-elsayed/ , Houda Bouamor https://aclanthology.org/people/houda-bouamor/unverified/ , Samhaa El-Beltagy https://aclanthology.org/people/samhaa-r-el-beltagy/ , Hend Al-Khalifa https://aclanthology.org/people/hend-al-khalifa/ , Ahmed Abdelali https://aclanthology.org/people/ahmed-abdelali/unverified/ , AbdelRahim Elmadany https://aclanthology.org/people/abdelrahim-elmadany/unverified/ , Abdellah El Mekki https://aclanthology.org/people/abdellah-el-mekki/ , Kareem Darwish https://aclanthology.org/people/kareem-darwish/ , Sakhar Alkhereyf https://aclanthology.org/people/sakhar-alkhereyf/ , Hamzah Luqman https://aclanthology.org/people/hamzah-luqman/ , Salam Khalifa https://aclanthology.org/people/salam-khalifa/ , Samar Magdy https://aclanthology.org/people/samar-magdy/unverified/ , Zaid Alyafeai https://aclanthology.org/people/zaid-alyafeai/ , Injy Hamed https://aclanthology.org/people/injy-hamed/unverified/ , Wajdi Zaghouani https://aclanthology.org/people/wajdi-zaghouani/ , Imed Zitouni https://aclanthology.org/people/imed-zitouni/unverified/ - Venue: - ArabicNLP https://aclanthology.org/venues/arabicnlp/ - SIG: - SIGARAB https://aclanthology.org/sigs/sigarab/ - Publisher: - Association for Computational Linguistics - Note: - Pages: - 1004–1009 - Language: - URL: - https://aclanthology.org/2026.arabicnlp-sharedtasks.123/ https://aclanthology.org/2026.arabicnlp-sharedtasks.123/ - DOI: - Cite ACL : - Mohammed Alsudairi, Naif Alharthi, Taif Nono, Faris Alasmary, Anas Adel Salamah, Shouq Sadah, and Lahouari Ghouti. 2026. Abjad AI at AraSeg Shared Task 2026: Auxiliary Loss for Arabic Sentence Segmentation https://aclanthology.org/2026.arabicnlp-sharedtasks.123/ . In Proceedings of The Fourth Arabic Natural Language Processing Conference: Shared Tasks , pages 1004–1009, Budapest, Hungary. Association for Computational Linguistics. - Cite Informal : - Abjad AI at AraSeg Shared Task 2026: Auxiliary Loss for Arabic Sentence Segmentation https://aclanthology.org/2026.arabicnlp-sharedtasks.123/ Alsudairi et al., ArabicNLP 2026 - PDF: - https://aclanthology.org/2026.arabicnlp-sharedtasks.123.pdf https://aclanthology.org/2026.arabicnlp-sharedtasks.123.pdf