Elamur Research⌕
☰
home / papers / 2025 / A Hybrid Transformer-Based Model for Sentiment Analysis of Arabic Dialect Hotel Reviews

A Hybrid Transformer-Based Model for Sentiment Analysis of Arabic Dialect Hotel Reviews

Paper · 2025 · Rawand Alfugaha and Mohammad AL-Smadi

Summary

The paper describes AraNLP, a system entered in the Ahasis shared task. The task asks for the overall polarity of a hotel review (Positive, Neutral or Negative), and the reviews are written in two under-resourced Arabic dialects, Saudi and Moroccan Darija. The organizers supplied a small dataset balanced across dialects and classes. The authors built a two-branch classifier. One branch is the [CLS] representation from the pre-trained AraELECTRA transformer, which carries contextual meaning. The other is a compact TF-IDF vector, which supplies surface lexical signals. The TF-IDF branch is projected through a linear layer with dropout and then joined with the transformer output before a classification head. Preprocessing was deliberately minimal, so dialect-specific spellings and words stay intact. The system reached 76% macro-F1, placing 5th, while the AraBERT baseline scored 56%. The error analysis shows that negation, contrastive "X but Y" reviews and implicit disappointment remain hard. Neutral reviews are also hard. The work suggests that cheap lexical features can still add value next to a transformer when dialect data is scarce.

Key points

  • The Ahasis data has 860 training and 216 test reviews. Both sets are roughly balanced across Saudi and Darija and across the three sentiment labels. No external data or cross-validation was used, and a small validation split was held out for early stopping.
  • The official metric is macro-averaged F1, so the neutral class counts as much as the other two.
  • The input is the AraELECTRA [CLS] embedding fused with a TF-IDF vector capped at 300 features. The authors tested 100 and 500 features, and 300 gave the best results.
  • The only preprocessing was tokenization, with no stemming, normalization, diacritic removal or stopword filtering, so dialect cues are preserved.
  • AraNLP scored 76% macro-F1 and ranked 5th. The organizers' AraBERT baseline scored 56% and ranked 13th.

Interesting ideas

  • I1. Leaving dialectal spellings unnormalized keeps variant forms that signal both region and sentiment, instead of folding them into a standard form too early.
  • I2. The sparse lexical branch is projected into its own small space before fusion, rather than being appended raw, which keeps the high-dimensional sparse input from overwhelming the dense input.
  • I3. The authors grouped their errors by linguistic phenomenon: negation, concessive contrast, faint praise such as "not bad", and implied letdown with no negative word.
  • I4. Neutral reviews are hard because one label covers factual reports, mixed opinions and off-topic text, so the authors propose a hierarchical decision instead of a flat one.
  • I5. Scoring with macro-F1 on class-balanced data stops a system from looking good by favouring frequent labels.

For Mova

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

  • I1 (tokens): The English tokenizer and lemmatizer could keep the original informal or regional surface form as a separate attribute next to the normalized lemma, so later stages can still see the variant.
  • I2 (tree): gparse could attach explicit symbolic cues, such as morphology flags or lexicon membership, through a learned projection, keeping these inputs inspectable while still letting the transformer use them.
  • I3 (frames): These categories map onto implicature and contrast handling in prag, and could become targeted test suites for concessive "but" clauses and expectation-violation statements.
  • I4 (frames): Speech-act classification could first separate factual statements from evaluative ones and only then assign polarity, so each step's reason stays explicit.
  • I5 (learn): Evaluations of rare frame types or speech acts could report per-class and macro scores, so gains on common categories cannot hide failures on rare ones.

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

🎓Qatar UniversityDoha · Qatar53 works · 1,027 citations

Also on the byline: Lusail University.

 Fields, topics and concepts

📚 Shelves 4The 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 works

🧭 Topics 2Research subjects and languages the work is about.Sentiment & emotion · 1,035 worksSemitic languages · 3,066 works

Cite

Cite as Rawand Alfugaha and Mohammad AL-Smadi. 2025. A Hybrid Transformer-Based Model for Sentiment Analysis of Arabic Dialect Hotel Reviews. Proceedings of the Shared Task on Sentiment Analysis for Arabic Dialects.

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{alfugaha2025hybrid,
  title = {{A Hybrid Transformer-Based Model for Sentiment Analysis of Arabic Dialect Hotel Reviews}},
  author = {Rawand Alfugaha and Mohammad AL-Smadi},
  booktitle = {Proceedings of the Shared Task on Sentiment Analysis for Arabic Dialects},
  year = {2025},
  doi = {10.26615/978-954-452-109-7-010},
  url = {https://doi.org/10.26615/978-954-452-109-7-010}
}

References

15 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Ahmed Abdelali, Sabit Hassan, Hamdy Mubarak, Kareem Darwish, and Younes Samih. 2021. Pre-training bert on arabic tweets: Practical considerations. arXiv preprint arXiv:2102.10684. ↗
  2. Wissam Antoun, Fady Baly, and Hazem Hajj. 2021. Araelectra: Pre-training text discriminators for arabic language understanding. In Proceedings of the Sixth Arabic Natural Language Processing Workshop, pages 191–195.
  3. Muhammad Abdul-Mageed, Abdelrahim Elmadany, et al. 2021. Arbert & marbert: Deep bidirectional transformers for arabic. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 7088–7105.
  4. Marouane Birjali, Mohammed Kasri, and Abderrahim Beni-Hssane. 2021. A comprehensive survey on sentiment analysis: Approaches, challenges and trends. Knowledge-Based Systems, 226:107134. find
  5. Ghizlane Bourahouat, Manar Abourezq, and Najima Daoudi. 2024. Improvement of moroccan dialect sentiment analysis using arabic bert-based models. J. Comput. Sci, 20(2):157–167. find
  6. Mohammad Al-Smadi, Mahmoud Al-Ayyoub, Yaser Jararweh, and Omar Qawasmeh. 2019. Enhancing aspect-based sentiment analysis of arabic hotels’ reviews using morphological, syntactic and semantic features. Information Processing & Management, 56(2):308–319. find
  7. Hanane Elfaik and El Habib Nfaoui. 2020. Deep bidirectional lstm network learning-based sentiment analysis for arabic text. Journal of Intelligent Systems, 30(1):395–412. find
  8. Mohammad AL-Smadi, Mahmoud M. Hammad, Sa’ad A. Al-Zboon, Saja AL-Tawalbeh, and Erik 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. arXiv preprint arXiv:2103.06678. ↗
  9. Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2019. Albert: A lite bert for self-supervised learning of language representations. arXiv preprint arXiv:1909.11942. ↗
  10. Omar Mansour, Eman Aboelela, Remon Talaat, and Mahmoud Bustami. 2025. Transformer-based ensemble model for dialectal arabic sentiment classification. PeerJ Computer Science, 11:e2644. find
  11. Abubakr H Ombabi, Wael Ouarda, and Adel M Alimi. 2020. Deep learning cnn–lstm framework for arabic sentiment analysis using textual information shared in social networks. Social Network Analysis and Mining, 10:1–13. find
  12. Maria Pontiki, Dimitrios Galanis, Haris Papageorgiou, Ion Androutsopoulos, Suresh Manandhar, Mohammad Al-Smadi, Mahmoud Al-Ayyoub, Yanyan Zhao, Bing Qin, Orphée De Clercq, et al. 2016. Semeval-2016 task 5: Aspect based sentiment analysis. In International workshop on semantic evaluation, pages 19–30.
  13. Mohammad Salameh, Houda Bouamor, and Nizar Habash. 2018. Fine-grained arabic dialect identification. In Proceedings of the 27th international conference on computational linguistics, pages 1332– 1344.
  14. Hager Saleh, Abdulaziz AlMohimeed, Rasha Hassan, Mandour M Ibrahim, Saeed Hamood Alsamhi, Moatamad Refaat Hassan, and Sherif Mostafa. 2025. Advancing arabic dialect detection with hybrid stacked transformer models. Frontiers in Human Neuroscience. find
  15. Bashar Talafha, Mohammad Ali, Muhy Eddin Za’ter, Haitham Seelawi, Ibraheem Tuffaha, Mostafa Samir, Wael Farhan, and Hussein Al-Natsheh. 2020. Multi-dialect arabic bert for country-level dialect identification. In Proceedings of the Fifth Arabic Natural Language Processing Workshop, pages 111–118.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/alfugaha-2025-hybrid-transformer-based-model-sentiment/