Elamur Research⌕
☰
home / papers / 2026 / An End-to-End Hybrid Framework for Rumour Detection in Low-Resources Algerian Dialect

An End-to-End Hybrid Framework for Rumour Detection in Low-Resources Algerian Dialect

Paper · 2026 · Dihia Lanasri and Fatima Benbarek

Summary

The paper addresses automatic rumour detection for Algerian social media posts about telecommunications. The setting is hard for three reasons: Darja is informal and mixes languages, it appears in both Arabic script and Latin-based Arabizi, and almost no labelled data exists. The authors built an end-to-end pipeline with three parts. The first assembles a corpus from Facebook posts, telecom messages generated with Google Gemini, and the FASSILA corpus remapped to rumour/non-rumour. Unlabelled items are annotated automatically by measuring semantic similarity to official operator statements. The second part transliterates the data into parallel Arabic-script and Arabizi versions. The third trains and compares classical ML, recurrent networks, transformers, LLMs and hybrids. The best system feeds transformer CLS embeddings into a simple classifier and reaches F1 = 0.84 on both scripts, above a re-implemented fine-tuned MarBERT baseline (0.82). The study also finds that pre-training on social-media data matters more than model size. Together, these results show that useful rumour detection is achievable for an under-resourced dialect.

Key points

  • The dataset combines three sources: real telecom posts (mostly from Facebook), synthetic dialect messages from Google Gemini, and FASSILA relabelled into a binary scheme. Labels for the unlabelled parts come from similarity to official facts collected from verified pages and the operator's website.
  • Hybrid models (CLS embeddings with SVM, Logistic Regression or Random Forest) achieve F1 = 0.84 on both the Arabic-script and Arabizi tracks. The choice of classifier makes almost no difference.
  • LLMs perform poorly on this task. Fine-tuned Llama3 reaches F1 = 0.72, and zero-shot prompting reaches only 0.43.
  • Script effects differ by model type:
    • Classical and transformer models differ by only about 1–2 F1 points between scripts.
    • Recurrent models do clearly better on Arabizi than on Arabic script.
  • DziriBERT and MarBERT miss rumours more often than they raise false alarms. Missed rumours are more frequent on the Arabic-script track, which the authors attribute to its morphological variability.

Interesting ideas

  • I1. Weak labelling against trusted references: a small set of trusted official statements can label a large unlabelled pool through semantic similarity, which is cheaper than manual annotation for a narrow domain.
  • I2. Parallel script tracks as an information check: the same data is converted into two parallel representations and models are trained on each. If scores are equal, the conversion lost little information.
  • I3. Simple-classifier probe on frozen embeddings: when SVM, LR and RF give nearly identical results on frozen embeddings, the representation itself holds the signal. This makes a quick diagnostic of encoder quality.
  • I4. Baselines framed as transferability tests: prior methods are re-run under one shared setup, and the comparison is presented as a test of how well they transfer rather than as a replication. This keeps the comparison fair and its claims honest.
  • I5. Error-direction analysis: the analysis checks which way errors lean (missed rumours versus false alarms) and links misses to paraphrase and surface variation. This points to where normalisation would help most.

For Mova

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

  • I1 (learn): Silver examples could be accepted only when they closely match a small hand-verified reference set, giving a cheap extra filter alongside witness-parser agreement.
  • I2 (tokens): Normalised and raw versions of noisy English (social-media spelling, dialect forms) could be tagged in parallel, and score parity would show whether the normaliser drops information.
  • I3 (tree): gparse's contextual vectors could be probed with several simple classifiers for UD relations, to check whether gains come from the encoder or from the decoding head.
  • I4 (learn): When older parsers or readers are benchmarked on Mova's own data, the reports should label results as transfer scores, kept separate from the originally published numbers.
  • I5 (frames): Claims in frames should be matched on lemmas and roles rather than surface strings, so reworded versions of the same who-did-what-to-whom are not missed.

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

Also on the byline: University of Algiers Benyoucef Benkhedda.

📚 Shelves 6 Computer Science · 9,086 works Artificial Intelligence · 8,889 works Arts and Humanities · 9,296 works Language and Linguistics · 9,206 works Social Sciences · 744 works Sociology and Political Science · 221 works

🧭 Topics 2 Semitic languages · 1,484 works Low-resource languages · 166 works

Cite

Cite as Dihia Lanasri and Fatima Benbarek. 2026. An End-to-End Hybrid Framework for Rumour Detection in Low-Resources Algerian Dialect. arXiv:2606.13411.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2606.13411 · arXiv 2606.13411

@misc{lanasri2026end,
  title = {{An End-to-End Hybrid Framework for Rumour Detection in Low-Resources Algerian Dialect}},
  author = {Dihia Lanasri and Fatima Benbarek},
  year = {2026},
  doi = {10.48550/arxiv.2606.13411},
  eprint = {2606.13411},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2606.13411}
}

References

18 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Carlos Castillo, Marcelo Mendoza, and Barbara Poblete. Information credibility on twitter. In Proceedings of the 20th International Conference on World Wide Web (WWW). ACM, 2011. find
  2. Vahed Qazvinian, Emily Rosengren, Dragomir R Radev, and Qiaozhu Mei. Rumor has it: Identifying misinformation in microblogs. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP). ACL, 2011. our card
  3. Arkaitz Zubiaga, Maria Liakata, and Rob Procter. Exploiting context for rumour detection in social media. In International Conference on Social Informatics, 2017. find
  4. Jing Ma, Wei Gao, and Kam-Fai Wong. Detecting rumors from microblogs with recurrent neural networks. In Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI), 2016. find
  5. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. Proceedings of NAACL-HLT, 2019. our card
  6. Xinyi Zhou and Reza Zafarani. A survey of fake news: Fundamental theories, detection methods, and opportunities. ACM Computing Surveys, 2020. ↗
  7. Mohamed Abdalla et al. Arabic natural language processing: A survey. ACM Computing Surveys, 2020. find
  8. K. Boukil et al. Arabic dialect processing: A survey and open challenges. In Proceedings of Arabic Language Processing Conference, 2021. find
  9. R. Alshalan, H. Al-Khalifa, et al. Detection of fake news in arabic content on social media. In IEEE Conference on Information and Communication Systems, 2020. find
  10. Chahnez Zakaria, Kamel Smaı̈li, Besma Sahnoun, Assia Chala, Radjaa Agagna, and Célia Amirat. Algerian arabizi rumour detection based on morphosyntactic analysis. International Journal of Knowledge Engineering and Data Mining, 8(1):43–66, 2023. ↗
  11. Amin Abdedaiem, Abdelhalim Hafedh Dahou, Mohamed Amine Cheragui, and Brigitte Mathiak. Fassila: a corpus for algerian dialect fake news detection and sentiment analysis. Procedia Computer Science, 244:397–407, 2024. find
  12. Arkaitz Zubiaga, Ahmet Aker, Kalina Bontcheva, Maria Liakata, and Rob Procter. Detection and resolution of rumours in social media: A survey. Acm Computing Surveys (Csur), 51(2):1–36, 2018. ↗
  13. Maysoon Alkhair, Karima Meftouh, Kamel Smaı̈li, and Nouha Othman. An arabic corpus of fake news: Collection, analysis and classification. In International Conference on Arabic Language Processing, pages 292–302. Springer, 2019. find
  14. Abdelouahab Hocini and Kamel Smaili. Detecting fake news: Exploring key features in multilingual arabic dialect corpus. In International Conference on Arabic Language Processing, pages 236–248. Springer, 2024. find
  15. Naelah O Bahurmuz, Ghada A Amoudi, Fatmah A Baothman, Amani T Jamal, Hanan S Alghamdi, and Areej M Alhothali. Arabic rumor detection using contextual deep bidirectional language modeling. IEEE Access, 10: 114907–114918, 2022. ↗
  16. Mohammed El Manar Righi, Djallel Eddine Boussahel, Djamila Mohdeb, Meriem Laifa, and Messaoud Bendiaf. Rumor stance classification: A case study on the propagation of political rumors on the algerian online social space. In 2022 International Conference on Advanced Aspects of Software Engineering (ICAASE), pages 1–6. find
  17. Mohamed Charafeddine Bousri, Riad Bensalem, Samah Bessa, Zineb Lamri, Chahnez Zakaria, and Nabila Bousbia. Rumor detection in algerian arabizi based on deep learning and associations. In International Symposium on Modelling and Implementation of Complex Systems, pages 165–176. Springer, 2022. find
  18. Khaoula Hamadouche, Kheira Zineb Bousmaha, Mohamed Yasine Bahi Amar, and Lamia Hadrich-Belguith. Detection of arabic and algerian fake news. Applied Computer Systems, 29(2):14–21, 2024. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/lanasri-2026-end-end-hybrid-framework-rumour/