Elamur Research⌕
☰
home / papers / 2026 / QalamID at AbjadAuthorID Shared Task: Morphology Matters, A Hybrid Ensemble for Arabic Authorship Attribution

QalamID at AbjadAuthorID Shared Task: Morphology Matters, A Hybrid Ensemble for Arabic Authorship Attribution

Paper · 2026 · Youssef Zaghloul

Summary

This paper presents the winning system of the AbjadAuthorID Shared Task, in which Arabic literary texts (classical, modern, and translated) must be assigned to one of 21 authors. The data pose three main difficulties. Class sizes are skewed at roughly 1:9, many documents exceed Transformer input limits, and genre conventions such as drama can mask individual style. The author combines five models through weighted soft voting:

  • calibrated LinearSVCs trained on character n-grams, word n-grams, and Stanza-derived Universal POS n-grams;
  • fine-tuned AraBERT v2 and XLM-RoBERTa, which process long texts as overlapping 512-token chunks pooled with learned attention.

Three additional techniques refine the ensemble. The weak POS model is kept but given a capped weight (the "Consultant Strategy"). Per-class probability multipliers, tuned on the validation confusion matrix, rebalance the output (the "Precision Scalpel"). Confident test predictions are then reused to retrain only the SVMs. On its own, the character n-gram SVM outperformed AraBERT, which suggests that Arabic authorial style is carried largely by morphological patterns. The result shows that inexpensive, morphology-aware features remain strong tools for Arabic stylometry.

Key points

  • Final result: The system ranked 1st, with a test macro F1 of 0.932 and accuracy of 0.963. Validation macro F1 was 0.948.
  • Single-model validation F1:
    • Character n-gram SVM (n = 2–5 within word boundaries, TF-IDF, 50k features): 0.92.
    • AraBERT v2: 0.87, with roughly 40× longer training time.
    • Word n-grams: 0.89.
    • POS n-grams: 0.67.
    • Classic stylometric features: 0.69.
  • Ensemble ablation (validation F1):
    • Uncalibrated ensemble: 0.936.
    • Adding pseudo-labeling (3,360 test samples with confidence ≥ 0.95): 0.941.
    • Adding the Precision Scalpel: 0.948.
  • Precision Scalpel effects:
    • Tharwat Abaza, often confused with translated Shakespeare, received a 2.0× boost. His recall rose from 0.47 to 0.68, an F1 gain of about 23 points.
    • Ahmed Shawqi received a 0.75× penalty, raising his F1 from 0.84 to 0.92.
  • Data and limitations:
    • Splits: 21 authors, 4,157 validation and 8,413 test samples, with duplicate training samples removed (about 6.5% of the training data).
    • The calibration depends on validation statistics and could overfit if the test distribution shifts.
    • The rarest author stays at about 0.88 F1.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Università Cattolica del Sacro CuoreMilan · Italy8 works · 33 citations🎓Universidad de ZaragozaZaragoza · Spain6 works · 5 citations

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksComputer Vision and Pattern Recognition · 3,486 works

🧭 Topics 2Research subjects and languages the work is about.Semitic languages · 3,066 worksMorphology · 420 works

Cite

Cite as Youssef Zaghloul. 2026. QalamID at AbjadAuthorID Shared Task: Morphology Matters, A Hybrid Ensemble for Arabic Authorship Attribution. Proceedings of the 2nd Workshop on NLP for Languages Using Arabic Script.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.abjadnlp-1.59

@inproceedings{zaghloul2026qalamid,
  title = {{QalamID at AbjadAuthorID Shared Task: Morphology Matters, A Hybrid Ensemble for Arabic Authorship Attribution}},
  author = {Youssef Zaghloul},
  booktitle = {Proceedings of the 2nd Workshop on NLP for Languages Using Arabic Script},
  year = {2026},
  doi = {10.18653/v1/2026.abjadnlp-1.59},
  url = {https://doi.org/10.18653/v1/2026.abjadnlp-1.59}
}

References

7 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Shadi Abudalfa, Saad Ezzini, Ahmed Abdelali, Hamza Alami, Abdessamad Benlahbib, Salmane Chafik, Mo El-Haj, Abdelkader El Mahdaouy, Mustafa Jarrar, Salima Lamsiyah, and Hamzah Luqman. 2025. The AraGenEval shared task on Arabic authorship style transfer and AI generated text detection. In Proceedings of The Third Arabic Natural Language Processing Conference at EMNLP 2025, Suzhou, China. Association for Computational Linguistics.
  2. Shadi Abudalfa, Saad Ezzini, Ahmed Abdelali, Mustafa Jarrar, Mo El-Haj, Nadir Durrani, Hassan Sajjad, Farah Adeeba, and Sina Ahmadi. 2026. Abjadauthorid: Authorship identification for arabic-script languages at abjadnlp 2026. In Proceedings of the 2nd Workshop on NLP for Languages Using Arabic Script (AbjadNLP 2026), co-located with the 19th Conference of the European Chapter of the Association for Computational Linguistics (EACL 2026), Rabat, Morocco. find
  3. Wissam Antoun, Fady Baly, and Hazem Hajj. 2020. Arabert: Transformer-based model for arabic language understanding. arXiv preprint arXiv:2003.00104. ↗
  4. John Burrows. 2002. ‘delta’: a measure of stylistic difference and a guide to likely authorship. Literary and linguistic computing, 17(3):267–287. find
  5. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186. ↗
  6. Peng Qi, Yuhao Zhang, Yuhui Zhang, Jason Bolton, and Christopher D Manning. 2020. Stanza: A python natural language processing toolkit for many human languages. arXiv preprint arXiv:2003.07082. ↗
  7. Efstathios Stamatatos. 2009. A survey of modern authorship attribution methods. Journal of the American Society for information Science and Technology, 60(3):538–556. A Branch Morphological (Char N-Gram) Lexical (Word N-Gram) Syntactic (POS N-Gram) Semantic (AraBERT v2) Multilingual (XLM-RoBERTa) Implementation Details & Configuration To ensure full replicability, we report the exact hyperparameters and configuration settings used for our final submission. A.1 Traditional Model Hyperparameters We utilized scikit-learn for feature extraction and classification. Table 3 details the specific configurations for the three traditional branches. Parameter N-Gram Range Analyzer Max Features Sublinear TF Classifier C Class Weight Calibration Char (2, 5) char_wb 50,000 True Word Syntactic (1, 3) (3, 5) word word 30,000 10,000 True True LinearSVC 1.0 1.0 0.5 ’balanced’ Platt Scaling (cv=3) *Table 4: Final voting weights. The morphological signal dominates (40%), while deep learning components contribute 30% combined.* Author Tharwat Abaza Kamel Kilani Ahmed Shawqi Others Mult. 2.00× 1.10× 0.75× 1.00× Rationale High Prec (0.94), Low Recall Mild boost for data scarcity "Gravity well" (high false pos) No adjustment *Table 5: Calibration multipliers ("Precision Scalpel"). Abaza required a significant boost to overcome genre-based confusion.* *Table 3: Hyperparameters for the three traditional branches. Note the lower regularization (C = 0.5) applied to the weaker syntactic branch.* A.2 Weight 0.40 0.20 0.10 0.15 0.15 find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zaghloul-2026-qalamid-abjadauthorid-shared-task-morphology/