Elamur Research⌕
☰
home / papers / 2024 / Functional Text Dimensions for Arabic Text Classification

Functional Text Dimensions for Arabic Text Classification

Paper · 2024 · Zeyd Ferhat, Abir Betka, Riyadh Barka et al.

Summary

Genre labels on web pages are often missing, vague or inconsistent between sources and annotators. This makes reliable genre classification hard, and Arabic lacks suitable evaluation data. The authors adapt Sharoff's Functional Text Dimensions (FTD) scheme to Arabic. In FTD, each category is defined by a diagnostic test question and a prototypical genre rather than by a source-specific label. They translated the category names and questions into Arabic. Eight annotators then searched the web manually for matching texts, filling one category at a time to a fixed quota. The result is the AFTD Corpus: 3,400 documents across 17 categories, released on Hugging Face. Classic machine-learning classifiers were benchmarked against fine-tuned BERT-style models. Arabic-specific models clearly beat the multilingual baseline, and CAMeLBERT reached an F1 of about 0.81. The work provides the first FTD resource for Arabic and a baseline for future genre-classification studies. Its main caveat is that no inter-annotator reliability was measured for this corpus.

Key points

  • AFTD Corpus: 17 FTD categories with 200 documents each (3,400 in total, 805,028 words); word counts per category are noticeably unbalanced.
  • Collection was done by 8 annotators (five men, three women) with at least some university education, using Arabic and English guidelines; each text kept its URL, collection date and source country.
  • Classic ML: logistic regression was best at 73% accuracy, and a decision tree was worst at 47%.
  • Fine-tuned transformers (512-token limit) scored F1 as follows: CAMeLBERT 81.43, MARBERT 80.62 (top accuracy at 81.68), QARiB 80.48, AraBERT 75.92, mBERT 70.75.
  • Main CAMeLBERT confusions: Academic predicted as Info 19 times and as Specialist 14 times, Fictive as Poetic 25 times, and Reporting as Report 13 times.

Interesting ideas

  • I1. Each category is pinned down by a short diagnostic question tied to a prototypical genre, which makes the labels easier to apply consistently across languages and annotators.
  • I2. FTD treats a text as a point in a multidimensional space of functions rather than forcing a single genre label, which leaves room for mixed texts.
  • I3. Data was gathered category by category with a fixed quota on a shared repository, and annotators moved on only once the target was met, guaranteeing balanced class counts.
  • I4. The confusion matrix was used to reveal which classes truly overlap (fiction vs poetry, academic vs specialist), not just to report a score.
  • I5. Comparing language-specific and multilingual pretrained models on the same benchmark shows how much native pretraining is worth for the task.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (frames): Speech-act and register labels in prag could each carry a test question plus a prototype example, so every assigned label can be justified by answering that question.
  • I2 (world): The world of a text could store graded function scores (narrative, instructive, argumentative), letting downstream steps know a fable also moralises.
  • I3 (learn): Silver-data and teacher-tier collection could use per-construction quotas so rare frames or syntactic patterns are not drowned by common ones.
  • I4 (learn): Evaluation could routinely list the top confused label pairs per stage to target repair rules or guideline clarifications at exactly those boundaries.
  • I5 (answer): Knowing the genre of an input text (tale, news, instruction) could route it to the matching reader, mirroring the benefit of specialised over generic models.

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

🎓University of ElouedEl Oued · Algeria2 works · 6 citations🎓University of BiskraBiskra · Algeria4 works · 7 citations

Also on the byline: University Mohamed Boudiaf of M'sila.

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksSocial Sciences · 1,772 worksGeneral Social Sciences · 227 works

🧭 Topics 1Research subjects and languages the work is about.Semitic languages · 2,908 works

Cite

Cite as Zeyd Ferhat, Abir Betka, Riyadh Barka, Zineddine Kahhoul, Selma Boutiba, Mohamed Tiar, Habiba Dahmani and Ahmed Abdelali. 2024. Functional Text Dimensions for Arabic Text Classification. Proceedings of the Second Arabic Natural Language Processing Conference.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2024.arabicnlp-1.29

@inproceedings{ferhat2024functional,
  title = {{Functional Text Dimensions for Arabic Text Classification}},
  author = {Zeyd Ferhat, Abir Betka, Riyadh Barka, Zineddine Kahhoul, Selma Boutiba, Mohamed Tiar, Habiba Dahmani and Ahmed Abdelali},
  booktitle = {Proceedings of the Second Arabic Natural Language Processing Conference},
  year = {2024},
  doi = {10.18653/v1/2024.arabicnlp-1.29},
  url = {https://doi.org/10.18653/v1/2024.arabicnlp-1.29}
}

References

22 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Melvil Dewey. 2004. A Classification and Subject Index for Cataloguing and Arranging the Books and Pamphlets of a Library. Project Gutenberg. find
  2. Ahmed Abdelali, Sabit Hassan, Hamdy Mubarak, Kareem Darwish, and Younes Samih. 2021. Pre-training bert on arabic tweets: Practical considerations. CoRR, abs/2102.10684. ↗
  3. Muhammad Abdul-Mageed, AbdelRahim A. Elmadany, and El Moatez Billah Nagoudi. 2021. Arbert & marbert: Deep bidirectional transformers for arabic. In ACL/IJCNLP (1), pages 7088–7105. Association for Computational Linguistics. ↗
  4. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. Cite arxiv:1810.04805. ↗
  5. H. Dittmann, J. Hardy, and L. Musgrave. 2007. Learn Library of Congress Classification. Library education series. TotalRecall Publications. find
  6. Omar Einea, Ashraf Elnagar, and Ridhwan Al Debsi. 2019. Sanad: Single-label arabic news articles dataset for automatic text categorization. Data in Brief, 25:104076. ↗
  7. Fumiyo Fukumoto and Yoshimi Suzuki. 2002. Manipulating large corpora for text classification. In EMNLP, pages 196–203. find
  8. Latifa Al-sulaiti and Eric Atwell. 2003. The design of a corpus of contemporary arabic. In University of Leeds School of Computing Research Series. ↗
  9. Sameh Alansary, Magdy H. Nagi, and Noha Adly. 2007. Building an international corpus of arabic ( ica ) : Progress of compilation stage. In 7th international conference on language engineering, Cairo, Egypt. find
  10. Wissam Antoun, Fady Baly, and Hazem M. Hajj. 2020. Arabert: Transformer-based model for arabic language understanding. CoRR, abs/2003.00104. ↗
  11. Go Inoue, Bashar Alhafni, Nurpeiis Baimukan, Houda Bouamor, and Nizar Habash. 2021. The interplay of variant, size, and task type in arabic pre-trained language models. CoRR, abs/2103.06678. ↗
  12. Klaus Krippendorff. 2006. Reliability in Content Analysis: Some Common Misconceptions and Recommendations. Human Communication Research, 30(3):411–433. find
  13. Robert Parker, David Graff, Ke Chen, Junbo Kong, and Kazuaki Maeda. 2011. Arabic gigaword fifth edition ldc2011t11. Philadelphia: Linguistic Data Consortium. Web Download. find
  14. Jorge Ramírez, Marcos Baez, Fabio Casati, and Boualem Benatallah. 2019. Crowdsourced dataset to study the generation and impact of text highlighting in classification tasks. BMC Research Notes, 12(820). ↗
  15. Motaz Saad and Wesam Ashour. 2010. Osac: Open source arabic corpora. In EEECS’10 the 6th International Symposium on Electrical and Electronics Engineering and Computer Science, pages 118–123. European University of Lefke, Cyprus. ↗
  16. Serge Sharoff. 2011. In the garden and in the jungle: Comparing genres in the bnc and internet. Genres on the web: Computational models and empirical studies, pages 149–166. find
  17. Serge Sharoff. 2018. Functional text dimensions for the annotation of web corpora. Corpora, 13(1):65–95. Firuza Urunbayevna. ↗
  18. Sobirova and Young Soon Cho. 2023. The Essence Of Multidimensional Analysis. European Scholar Journal, 4(12):31–32. find
  19. Vít Suchomel and Jan Kraus. 2022. Semi-manual annotation of topics and genres in web corpora, the cheap and fast way. In RASLAN, pages 141–148. Tribun EU. find
  20. Keyu Yang, Yunjun Gao, Lei Liang, Song Bian, Lu Chen, and Baihua Zheng. 2021. Crowdtc: Crowd-powered learning for text classification. ACM Trans. Knowl. Discov. Data, 16(1). ↗
  21. Petr Zemánek. 2001. Clara (corpus linguae arabicae): An overview. In EMNLP. find
  22. Imad Zeroual, Dirk Goldhahn, Thomas Eckart, and Abdelhak Lakhouaja. 2019. Osian: Open source international arabic news corpus - preparation and integration into the clarin-infrastructure. In WANLP@ACL 2019, pages 175–182. Association for Computational Linguistics. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/ferhat-2024-functional-text-dimensions-arabic-text/