---
title: "Quasar at MAHED Shared Task : Decoding Emotions and"
description: "The paper describes team Quasar's systems for the MAHED 2025 Shared Task at ArabicNLP 2025. 21 references, cited by 1 paper here."
authors: ["Md Sagor Chowdhury", "Adiba Fairooz Chowdhury"]
year: 2025
venue: "Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks"
source: "https://aclanthology.org/2025.arabicnlp-sharedtasks.95/"
doi: "10.18653/v1/2025.arabicnlp-sharedtasks.95"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/chowdhury-2025-quasar-mahed-shared-task-decoding/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/chowdhury-2025-quasar-mahed-shared-task-decoding/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2025](https://md.research.elamur.ai/free/en/papers/2025/page/15/) / Quasar at MAHED Shared Task : Decoding Emotions and Offense in Arabic Text using LLM and Transformer-Based Approaches

# Quasar at MAHED Shared Task : Decoding Emotions and Offense in Arabic Text using LLM and Transformer-Based Approaches

**Authors:** [Md Sagor Chowdhury](https://md.research.elamur.ai/free/en/authors/md-sagor-chowdhury/) and [Adiba Fairooz Chowdhury](https://md.research.elamur.ai/free/en/authors/adiba-fairooz-chowdhury/) · **Year:** 2025 · **Venue:** [Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks](https://md.research.elamur.ai/free/en/venues/arabicnlp/) ([ArabicNLP 2025](https://md.research.elamur.ai/free/en/venues/arabicnlp/2025/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2025.arabicnlp-sharedtasks.95/) · [PDF](https://aclanthology.org/2025.arabicnlp-sharedtasks.95.pdf) · [DOI 10.18653/v1/2025.arabicnlp-sharedtasks.95](https://doi.org/10.18653/v1/2025.arabicnlp-sharedtasks.95) · [OpenAlex](https://openalex.org/W4416037625)

**Citations:** 0 citations (OpenAlex, 2026-10-14) · [1](https://md.research.elamur.ai/free/en/papers/chowdhury-2025-quasar-mahed-shared-task-decoding/cited-by/) in our library, of the 149,439 papers whose bibliographies we read

## Summary
The paper describes team Quasar's systems for the MAHED 2025 Shared Task at ArabicNLP 2025. The task asks for automatic detection of harmful and positive speech in Arabic social media. Subtask 1 sorts posts into hate, hope or not_applicable. Subtask 2 is a multi-label problem covering emotion, offensiveness and directed hate, built on the EmoHopeSpeech dataset. The training data for Subtask 1 was skewed, so the authors balanced the classes in three ways: they added external corpora, applied Arabic synonym substitution, and used a back-translation loop through English with WordNet/spaCy synonym swaps in the middle. They then compared multilingual and Arabic-specific encoders (XLM-RoBERTa-Large, AraBERTv2, asafaya BERT, Davlan XLM-R) with LLMs fitted with classification heads (Gemma-7B, Qwen2.5-14B-Instruct), across many data configurations. Gemma-7B with DoRA adaptation did best on Subtask 1 at about 0.67 macro-F1, which ranked 12th. On Subtask 2, several models tied at 0.48 macro-F1, which ranked 6th. The experiments also show that the choice of preprocessing and balancing can shift results more than the choice of model.

## Key points
- The Subtask 1 training set had 6,890 posts: 3,697 not_applicable, 1,892 hope and 1,301 hate. After augmentation it was balanced at 3,697 per class, 11,091 in total.
- Hope augmentation drew on three sources:
  - 130 posts from an external Arabic optimism dataset;
  - 1,675 posts made by dictionary-based synonym replacement;
  - a pass of back-translation via Google Translate.
- Gemma-7B with a DoRA configuration scored 0.67 macro-F1 on an unprocessed test set and 0.64 on a preprocessed one. In contrast, XLM-RoBERTa-Large trained on the augmented and preprocessed data dropped to 0.33 because its predictions leaned heavily toward hope.
- On Subtask 2 (EmoHopeSpeech, 5,960 rows), Gemma-7B, aubmindlab/bert-base-arabertv2 and Qwen2.5-14B-Instruct all reached 0.48 macro-F1. An ensemble of XLM-RoBERTa, Gemma and dehatebert scored lower, at 0.43.
- The specialised hate models (hossam87/bert-base-arabic-hate-speech and dehatebert-mono-arabic) showed a strong bias toward the hate label.

## Interesting ideas
- Synthetic oversampling can backfire: the augmented set pushed one model toward the inflated class, so checking the per-class prediction distribution is a cheap way to catch this.
- The authors treated test-time preprocessing as its own experimental factor (processed vs. unprocessed test set) and found that matching or mismatching it changed scores.
- Their back-translation recipe changes wording in the pivot language, swapping synonyms only for content-word parts of speech, so it produces paraphrases that keep the meaning.
- They analysed Subtask 2 per component and saw that directed hate is a harder decision nested inside offensiveness, because the line between general offense and targeted hate is subtle.
- The error analysis found that religiously phrased hopeful messages were often read as neutral or humorous, a failure tied to register.

## For Mova

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

- **I1** (learn): Any class-rebalanced silver data should be paired with a negative control comparing label frequencies before and after training, so that augmentation-induced bias is caught early.
- **I2** (tokens): Normalization rules in crate en could be checked by running the same evaluation on raw and normalized input, which would reveal where cleanup removes useful signal.
- **I3** (back): A similar constrained paraphrase loop could generate test pairs for checking that the text → graph → text compressor maps meaning-equivalent sentences to the same graph.
- **I4** (frames): Directed hate could be modelled as an offensive speech act that has an explicit target role in the frame, so the who-is-attacked slot makes the distinction explainable.
- **I5** (frames): Irony and implicature detection in prag should take the text's register or genre into account, so that formulaic devout language is not mistaken for sarcasm.

*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

- [Chittagong University of Engineering & Technology](https://md.research.elamur.ai/free/en/universities/chittagong-university-of-engineering-technology/) — University, Chittagong, Bangladesh · 25 works · 29 citations

📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 16,039 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 15,891 works

🧭 Topics (2): [😊 Sentiment & emotion](https://md.research.elamur.ai/free/en/topics/sentiment-emotion/) · 972 works · [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) · 2,908 works

## Cite

Cite as: Md Sagor Chowdhury and Adiba Fairooz Chowdhury. 2025. Quasar at MAHED Shared Task : Decoding Emotions and Offense in Arabic Text using LLM and Transformer-Based Approaches. Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks.

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

```bibtex
@inproceedings{chowdhury2025quasar,
  title = {{Quasar at MAHED Shared Task : Decoding Emotions and Offense in Arabic Text using LLM and Transformer-Based Approaches}},
  author = {Md Sagor Chowdhury and Adiba Fairooz Chowdhury},
  booktitle = {Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks},
  year = {2025},
  doi = {10.18653/v1/2025.arabicnlp-sharedtasks.95},
  url = {https://doi.org/10.18653/v1/2025.arabicnlp-sharedtasks.95}
}
```

- [← previous Qais at BAREC Shared Task 2025: A Fine-Grained Approach for Arabic Readability Classification Using a pre-trained model](https://md.research.elamur.ai/free/en/papers/ahmad-2025-qais-barec-shared-task-2025/)
- [next → Rationale-Guided Distillation for E-Commerce Relevance Classification: Bridging Large Language Models and Lightweight Cross-Encoders](https://md.research.elamur.ai/free/en/papers/agrawal-2025-rationale-guided-distillation-e-commerce/)
- [↑ back to papers of 2025](https://md.research.elamur.ai/free/en/papers/2025/page/15/)

## References

21 references, 10 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Firoj Alam, Md Rafiul Biswas, Uzair Shah, Wajdi Zaghouani, and Georgios Mikros. 2024a. [Propaganda to hate: A multimodal analysis of arabic memes with multi-agent llms](https://md.research.elamur.ai/free/en/papers/alam-2024-propaganda-hate-multimodal-analysis-arabic/). In International Conference on Web Information Systems Engineering, pages 380– 390. Springer. Firoj Alam, Abul Hasnat, Fatema Ahmad, Md.
2. Arid Hasan, and Maram Hasanain. 2024b. [ArMeme: Propagandistic content in Arabic memes](https://md.research.elamur.ai/free/en/papers/alam-2024-armeme-propagandistic-content-arabic-memes/). In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 21071– 21090, Miami, Florida, USA. Association for Computational Linguistics.
3. Firoj Alam, Hamdy Mubarak, Wajdi Zaghouani, Giovanni Da San Martino, and Preslav Nakov. 2022. [Overview of the WANLP 2022 shared task on propaganda detection in Arabic](https://md.research.elamur.ai/free/en/papers/alam-2022-overview-wanlp-2022-shared-task/). In Proceedings of the The Seventh WANLP, Abu Dhabi, United Arab Emirates (Hybrid). ACL.
4. Sai Saket Aluru, Binny Mathew, Punyajoy Saha, and Animesh Mukherjee. 2020. Deep learning models for multilingual hate speech detection. arXiv preprint arXiv:2004.06465. [↗](https://arxiv.org/abs/2004.06465)
5. Hossam. 2023. Bert base arabic hate speech detection model. Hugging Face model repository. [find](https://scholar.google.com/scholar?q=Hossam%2E+2023%2E+Bert+base+arabic+hate+speech+detection+model%2E+Hugging+Face+model+repository%2E)
6. Wissam Antoun, Fady Baly, and Hazem Hajj. 2020. AraBERT: A pre-trained arabic language model. In Proceedings of the 4th Workshop on Open-Source Arabic Corpora and Processing Tools, Marseille, France. European Language Resource Association. [find](https://scholar.google.com/scholar?q=Wissam+Antoun%2C+Fady+Baly%2C+and+Hazem+Hajj%2E+2020%2E+AraBERT%3A+A+pre%2Dtrained+arabic+language+model%2E+In+Proceedings+of+the+4th+Workshop+on+Open%2DSource+Arabic+Corpora+and+Processing+Tools%2C+Marseille%2C+France%2E+)
7. Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, et al. 2024. Gemma: Open models based on gemini research and technology. arXiv preprint arXiv:2403.08295. Md. [↗](https://arxiv.org/abs/2403.08295)
8. Rafiul Biswas and Wajdi Zaghouani. 2025a. [An annotated corpus of arabic tweets for hate speech analysis](https://md.research.elamur.ai/free/en/papers/zaghouani-2025-annotated-corpus-arabic-tweets-hate/). CoRR, abs/2505.11969. Md. [↗](https://arxiv.org/abs/2505.11969)
9. Rafiul Biswas and Wajdi Zaghouani. 2025b. [Emohopespeech: An annotated dataset of emotions and hope speech in english and arabic](https://md.research.elamur.ai/free/en/papers/zaghouani-2025-emohopespeech-annotated-dataset-emotions-hope/). CoRR, abs/2505.11959. [↗](https://arxiv.org/abs/2505.11959)
10. Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, Florence, Italy. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Alexis+Conneau%2C+Kartikay+Khandelwal%2C+Naman+Goyal%2C+Vishrav+Chaudhary%2C+Guillaume+Wenzek%2C+Francisco+Guzm%C3%A1n%2C+Edouard+Grave%2C+Myle+Ott%2C+Luke+Zettlemoyer%2C+and+Veselin+Stoyanov%2E+2020%2E+Unsupervised+cross%2Dlingu)
11. Davlan Team. 2023. Xlm-roberta base fine-tuned for arabic. No specific academic paper is associated with Davlan/xlm-roberta-base-finetuned-arabic; refer to the Hugging Face model page for details. [find](https://scholar.google.com/scholar?q=Davlan+Team%2E+2023%2E+Xlm%2Droberta+base+fine%2Dtuned+for+arabic%2E+No+specific+academic+paper+is+associated+with+Davlan%2Fxlm%2Droberta%2Dbase%2Dfinetuned%2Darabic%3B+refer+to+the+Hugging+Face+model+page+for+details%2E)
12. Dimitar Dimitrov, Firoj Alam, Maram Hasanain, Abul Hasnat, Fabrizio Silvestri, Preslav Nakov, and Giovanni Da San Martino. 2024. Semeval-2024 task 4: Multilingual detection of persuasion techniques in memes. In Proceedings of the 18th International Workshop on Semantic Evaluation (SemEval-2024), Mexico City, Mexico. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Dimitar+Dimitrov%2C+Firoj+Alam%2C+Maram+Hasanain%2C+Abul+Hasnat%2C+Fabrizio+Silvestri%2C+Preslav+Nakov%2C+and+Giovanni+Da+San+Martino%2E+2024%2E+Semeval%2D2024+task+4%3A+Multilingual+detection+of+persuasion+techniques+in)
13. Maram Hasanain, Fatema Ahmad, and Firoj Alam. 2024a. Large language models for propaganda span annotation. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 14522– 14532, Miami, Florida, USA. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Maram+Hasanain%2C+Fatema+Ahmad%2C+and+Firoj+Alam%2E+2024a%2E+Large+language+models+for+propaganda+span+annotation%2E+In+Findings+of+the+Association+for+Computational+Linguistics%3A+EMNLP+2024%2C+pages+14522%E2%80%93+14532%2C)
14. Maram Hasanain, Fatema Ahmed, and Firoj Alam. 2024b. Can gpt-4 identify propaganda? annotation and detection of propaganda spans in news articles. In Proceedings of the 2024 Joint International Conference On Computational Linguistics, Language Resources And Evaluation, LREC-COLING 2024, Torino, Italy. [find](https://scholar.google.com/scholar?q=Maram+Hasanain%2C+Fatema+Ahmed%2C+and+Firoj+Alam%2E+2024b%2E+Can+gpt%2D4+identify+propaganda%3F+annotation+and+detection+of+propaganda+spans+in+news+articles%2E+In+Proceedings+of+the+2024+Joint+International+Confer)
15. Maram Hasanain, Firoj Alam, Hamdy Mubarak, Samir Abdaljalil, Wajdi Zaghouani, Preslav Nakov, Giovanni Da San Martino, and Abed Alhakim Freihat. 2023. [ArAIEval Shared Task: Persuasion techniques and disinformation detection in arabic text](https://md.research.elamur.ai/free/en/papers/hasanain-2023-araieval-shared-task-persuasion-techniques/). In Proceedings of the First Arabic Natural Language Processing Conference (ArabicNLP 2023), Singapore. Association for Computational Linguistics.
16. Ali Safaya, Moutasem Abdullatif, and Deniz Yuret. 2020. KUISAIL at SemEval-2020 task 12: BERT-based multi-label classification for offensive language detection. In Proceedings of the Fourteenth Workshop on Semantic Evaluation, Barcelona (online). International Committee for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Ali+Safaya%2C+Moutasem+Abdullatif%2C+and+Deniz+Yuret%2E+2020%2E+KUISAIL+at+SemEval%2D2020+task+12%3A+BERT%2Dbased+multi%2Dlabel+classification+for+offensive+language+detection%2E+In+Proceedings+of+the+Fourteenth+Worksh)
17. An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, et al. 2024. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115. [↗](https://arxiv.org/abs/2412.15115)
18. Wajdi Zaghouani and Md Rafiul Biswas. 2025. [Emohopespeech: An annotated dataset of emotions and hope speech in english and arabic](https://md.research.elamur.ai/free/en/papers/zaghouani-2025-emohopespeech-annotated-dataset-emotions-hope/). arXiv preprint arXiv:2505.11959. [↗](https://arxiv.org/abs/2505.11959)
19. Wajdi Zaghouani, Md Rafiul Biswas, Mabrouka Bessghaier, Shimaa Ibrahim, Georgios Mikros, Abul Hasnat, and Firoj Alam. 2025. [MAHED shared task: Multimodal detection of hope and hate emotions in arabic content](https://md.research.elamur.ai/free/en/papers/zaghouani-2025-mahed-shared-task-multimodal-detection/). In Proceedings of the Third Arabic Natural Language Processing Conference (ArabicNLP 2025), Suzhou, China. Association for Computational Linguistics. Wajdi Zaghouani, Hamdy Mubarak, and Md.
20. Rafiul Biswas. 2024a. [So hateful! building a multi-label hate speech annotated Arabic dataset](https://md.research.elamur.ai/free/en/papers/zaghouani-2024-so-hateful-building-multi-label/). In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 15044–15055, Torino, Italia. ELRA and ICCL.
21. Wajdi Zaghouani, Hamdy Mubarak, and Md Rafiul Biswas. 2024b. [So hateful! building a multi-label hate speech annotated arabic dataset](https://md.research.elamur.ai/free/en/papers/zaghouani-2024-so-hateful-building-multi-label/). In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 15044–15055.

*Updated 31 October 2026*