--- title: "ANLP-UniSo at MAHED Shared Task: Detection of Hate and" description: "The paper presents the system a University of Sousse team submitted for the first… 0.73 precision. 4 references, cited by 2 papers here." authors: ["Yasmine El Abed", "Mariem Ben Arbia", "Saoussen Ben Chaabene", "Omar Trigui"] year: 2025 venue: "Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks" source: "https://aclanthology.org/2025.arabicnlp-sharedtasks.87/" doi: "10.18653/v1/2025.arabicnlp-sharedtasks.87" 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/abed-2025-anlp-uniso-mahed-shared-task/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/abed-2025-anlp-uniso-mahed-shared-task/) [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/13/) / ANLP-UniSo at MAHED Shared Task: Detection of Hate and Hope Speech in Arabic Social Media based on XLM-RoBERTa and Logistic Regression # ANLP-UniSo at MAHED Shared Task: Detection of Hate and Hope Speech in Arabic Social Media based on XLM-RoBERTa and Logistic Regression **Authors:** Yasmine El Abed, Mariem Ben Arbia, Saoussen Ben Chaabene and [Omar Trigui](https://md.research.elamur.ai/free/en/authors/omar-trigui/) · **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.87/) · [PDF](https://aclanthology.org/2025.arabicnlp-sharedtasks.87.pdf) · [DOI 10.18653/v1/2025.arabicnlp-sharedtasks.87](https://doi.org/10.18653/v1/2025.arabicnlp-sharedtasks.87) · [OpenAlex](https://openalex.org/W4416037754) **Citations:** 0 citations (OpenAlex, 2026-10-14) · [2](https://md.research.elamur.ai/free/en/papers/abed-2025-anlp-uniso-mahed-shared-task/cited-by/) in our library, of the 149,439 papers whose bibliographies we read ## Summary The paper presents the system a University of Sousse team submitted for the first subtask of MAHED 2025's shared evaluation. The task is to label short Arabic social-media posts, written in MSA or dialect, as Hate, Hope, or not_applicable. The pipeline has four main parts: - **Cleaning:** removing diacritics, URLs, emojis, letter elongations, stop words and rare tokens. - **Augmentation:** back-translating Hope sentences through English to add more Hope examples. - **Features:** extracting XLM-RoBERTa embeddings, then applying SMOTE in that embedding space to rebalance the classes. - **Classifiers:** Logistic Regression and an LSTM, both trained on the rebalanced features. The abstract also mentions a zero-shot similarity check on predictions, but the paper gives no details. The training data has 6,890 sentences. Evaluation used two phases: an initial validation set, then an updated validation set plus a hidden test set. The two classifiers score close to each other, with macro F1 around 0.56–0.57. The LSTM finds more Hate and Hope posts but makes more false alarms. LR was steadier across phases, and its official leaderboard score was 0.60. The work matters because it treats detecting hope as seriously as detecting hate, which is a rare focus for Arabic. Its error analysis also points to sarcasm and misleading positive words as the main obstacles. ## Key points - **Training set:** 6,890 sentences, with not_applicable 3,697, Hope 1,892 and Hate 1,301. The paper's text calls hope and not_applicable the minority classes, which does not match this table. - **Macro F1 on development:** LR 0.5658, LSTM 0.5561. - **Macro F1 on test:** LR 0.5740, LSTM 0.5551. The official leaderboard reports 0.60 for the final LR-based submission. - **LSTM on test:** strong recall for Hate (0.74) and Hope (0.70), but low precision for both. For not_applicable it reaches 0.73 precision and only 0.42 recall. - **Related resources:** the MAHED data builds on So Hateful!, Emo-HopeSpeech and an annotated corpus of Arabic hate-speech tweets. The authors plan to try MARBERT and few-shot methods next. ## Interesting ideas - A plain logistic regression on frozen contextual embeddings matched the LSTM and was more stable across evaluation phases, so a transparent linear probe is a strong, honest baseline before adding heavier models. - Reporting per-class precision/recall instead of just macro F1 showed the two models fail in opposite directions: one under-predicts the rare classes, the other over-predicts them. - Scoring the system twice, first on the initial validation set and then on revised validation and test sets, revealed which classifier kept its behaviour when the data shifted. - The error analysis found that sarcasm, figurative wording and individual positive words can flip a post's predicted stance against its overall meaning, which is why a sentence-level bag of features breaks down. - Back-translation through a pivot language was used to create varied paraphrases of an under-represented class without new annotation. ## For Mova How these ideas could serve Mova, our small open language model — its files, stages and tests: - **I1** (learn): Add a linear probe on gparse encoder states as a fixed baseline, so any gain claimed by a more complex component is measured against it. - **I2** (learn): Evaluation reports for frame or speech-act labelers should show each class's precision and recall, so conservative and trigger-happy failure modes stay visible. - **I3** (learn): Score silver-trained parsers against at least two independently built dev slices and treat a large disagreement between them as a warning sign. - **I4** (frames): The irony and implicature layer could compare the polarity of local words with the role structure of the whole frame (who is hoped for and who is attacked) and flag mismatches as possible sarcasm. - **I5** (back): The text → graph → text round trip could produce paraphrases of rare constructions, and keeping only those whose re-parse yields the same graph would give checked extra training data. *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 Sousse](https://md.research.elamur.ai/free/en/universities/university-of-sousse/) — University, Sousse, Tunisia · 4 works · 19 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 (1): [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) · 2,908 works ## Cite Cite as: Yasmine El Abed, Mariem Ben Arbia, Saoussen Ben Chaabene and Omar Trigui. 2025. ANLP-UniSo at MAHED Shared Task: Detection of Hate and Hope Speech in Arabic Social Media based on XLM-RoBERTa and Logistic Regression. 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{abed2025anlp, title = {{ANLP-UniSo at MAHED Shared Task: Detection of Hate and Hope Speech in Arabic Social Media based on XLM-RoBERTa and Logistic Regression}}, author = {Yasmine {El Abed}, Mariem Ben Arbia, Saoussen Ben Chaabene and Omar Trigui}, booktitle = {Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks}, year = {2025}, doi = {10.18653/v1/2025.arabicnlp-sharedtasks.87}, url = {https://doi.org/10.18653/v1/2025.arabicnlp-sharedtasks.87} } ``` - [← previous AmaSQuAD: A Benchmark for Amharic Extractive Question Answering](https://md.research.elamur.ai/free/en/papers/hailemariam-2025-amasquad-benchmark-amharic-extractive-question/) - [next → ANPLers at IqraEval Shared task: Adapting Whisper-large-v3 as Speech-to-Phoneme for Qur'anic Recitation Mispronunciation Detection](https://md.research.elamur.ai/free/en/papers/qandos-2025-anplers-iqraeval-shared-task-adapting/) - [↑ back to papers of 2025](https://md.research.elamur.ai/free/en/papers/2025/page/13/) ## References 4 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. Ahmad, M., Usman, S., Farid, H., Ameer, I., Muzammil, M., Hamza, A., Sidorov, G., and Batyrshin, I. (2024). Hope speech detection using social media discourse (posi-vox-2024): A transfer learning approach. Journal of Language and Education, 10(4):31–43. [↗](https://doi.org/10.17323/jle.2024.22443) 2. Antoun, W., Baly, F., and Hajj, H. (2020). [AraBERT: Transformer-based model for arabic language understanding](https://md.research.elamur.ai/free/en/papers/antoun-2020-arabert-transformer-based-model-arabic/). arXiv preprint arXiv:2003.00104. [↗](https://arxiv.org/abs/2003.00104) 3. Chakravarthi, B. R., Muralidaran, V., Hande, A., Philip, J., McCrae, J. P., Buitelaar, P., Ponnusamy, R., Suryawanshi, S., Sherly, E., and Bandyopadhyay, S. (2022). Hope speech detection for equality, diversity and inclusion. In Proceedings of the Second Workshop on Language Technology for Equality, Diversity and Inclusion (LT-EDI 2022), pages 378–388, Dublin, Ireland. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Chakravarthi%2C+B%2E+R%2E%2C+Muralidaran%2C+V%2E%2C+Hande%2C+A%2E%2C+Philip%2C+J%2E%2C+McCrae%2C+J%2E+P%2E%2C+Buitelaar%2C+P%2E%2C+Ponnusamy%2C+R%2E%2C+Suryawanshi%2C+S%2E%2C+Sherly%2C+E%2E%2C+and+Bandyopadhyay%2C+S%2E+%282022%29%2E+Hope+speech+detection+for+equality%2C) 4. Zaghouani, W., Mubarak, H., and Biswas, M. R. (2024). [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 1 November 2026*