--- title: "MarsadLab at AraHealthQA — paper by Bessghaier et al. (2025)" description: "This paper describes the MarsadLab entry for Track 1 (MentalQA) of the AraHealthQA 2025 Shared Task. 15 references, cited by 1 paper here." authors: ["Mabrouka Bessghaier", "Shimaa Ibrahim", "Md. Rafiul Biswas", "Wajdi Zaghouani"] year: 2025 venue: "Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks" source: "https://aclanthology.org/2025.arabicnlp-sharedtasks.33/" doi: "10.18653/v1/2025.arabicnlp-sharedtasks.33" 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/bessghaier-2025-marsadlab-arahealthqa-hybrid-contextual-lexical/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/bessghaier-2025-marsadlab-arahealthqa-hybrid-contextual-lexical/) [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/) / MarsadLab at AraHealthQA: Hybrid Contextual–Lexical Fusion with AraBERT for Question and Answer Categorization # MarsadLab at AraHealthQA: Hybrid Contextual–Lexical Fusion with AraBERT for Question and Answer Categorization **Authors:** [Mabrouka Bessghaier](https://md.research.elamur.ai/free/en/authors/mabrouka-bessghaier/), [Shimaa Ibrahim](https://md.research.elamur.ai/free/en/authors/shimaa-amer-ibrahim/), [Md. Rafiul Biswas](https://md.research.elamur.ai/free/en/authors/md-rafiul-biswas/) and [Wajdi Zaghouani](https://md.research.elamur.ai/free/en/authors/wajdi-zaghouani/) · **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.33/) · [PDF](https://aclanthology.org/2025.arabicnlp-sharedtasks.33.pdf) · [DOI 10.18653/v1/2025.arabicnlp-sharedtasks.33](https://doi.org/10.18653/v1/2025.arabicnlp-sharedtasks.33) · [OpenAlex](https://openalex.org/W4416037820) **Citations:** 0 citations (OpenAlex, 2026-10-14) · [1](https://md.research.elamur.ai/free/en/papers/bessghaier-2025-marsadlab-arahealthqa-hybrid-contextual-lexical/cited-by/) in our library, of the 149,439 papers whose bibliographies we read ## Summary This paper describes the MarsadLab entry for Track 1 (MentalQA) of the AraHealthQA 2025 Shared Task. The track has two multi-label problems on Arabic mental-health text. Subtask A assigns questions to topical categories, such as Diagnosis, Treatment or Epidemiology, plus a default "Other" class. Subtask B labels answers by their strategy: Information, Direct Guidance or Emotional Support. The data is a small set of roughly 500 hand-annotated question–answer pairs, mostly in Modern Standard Arabic with some dialect, collected from online user posts. The system has two branches. One uses AraBERTv2 sentence embeddings. The other is a block of handcrafted binary indicators: class lexicons mined automatically from the training data, and keyword lists written by hand to capture pragmatic markers. The two branches are concatenated and fed to a linear sigmoid classifier. On the official test set the system reached weighted F1 0.55 for Task A and 0.79 for Task B. Ablations show that the keyword cues gave a modest gain only for question classification. The authors say the hybrid design matters because it keeps part of the decision interpretable when training data is scarce. ## Key points - **Backbone and fusion:** aubmindlab/bert-base-arabertv2 produces a 768-dimensional pooled vector. This is joined with binary feature vectors, passed through dropout and a linear head, and thresholded at 0.5. If no label passes the threshold, an argmax fallback assigns one. - **Preprocessing:** text is normalized (diacritics removed, alif/ya/taa marbuta variants unified), stopwords are filtered, and words are reduced with the ISRI stemmer. - **Lexicon features:** each class lexicon holds the 40 most frequent stemmed non-stopword tokens for that class in training. The feature only records whether any lexicon token appears; repeated matches do not raise its value. - **Task A ablation (F1/Jaccard):** | Configuration | F1 | Jaccard | |---|---|---| | AraBERT only | 0.52 | 0.39 | | +Lexicon | 0.51 | 0.37 | | +Keywords | 0.56 | 0.43 | | Both (submitted) | 0.55 | 0.41 | | SVM on features only | 0.45 | 0.34 | - **Task B ablation (F1/Jaccard):** | Configuration | F1 | Jaccard | |---|---|---| | AraBERT only | 0.79 | 0.69 | | +Lexicon | 0.77 | 0.67 | | +Keywords | 0.78 | 0.68 | | Both | 0.79 | 0.67 | | SVM on features only | 0.74 | 0.62 | Emotional Support was often confused with guidance labels. ## Interesting ideas - An argmax fallback guarantees that a multi-label classifier never returns an empty label set, which avoids silent "no answer" outputs on low-confidence inputs - Answer types are separated using hand-picked pragmatic markers (definitional phrases, directive verbs, comforting expressions), which helped mainly on the pragmatic categories - Class lexicons are mined automatically as the top frequent stems per label and used only as presence flags, which gives cheap, readable evidence for each decision - Features are also tested alone with an SVM, so the authors could see that the cues carried real signal even though adding them to AraBERT gave little gain - The authors read the flat results honestly: they blame naive concatenation, not the features, when a strong encoder swamps sparse binary signals ## For Mova How these ideas could serve Mova, our small open language model — its files, stages and tests: - **I1** (frames): speech-act tagging of an utterance could use the same rule, so that every clause always gets at least its most likely act and the chosen act stays traceable. - **I2** (frames): marker lists like these could act as explicit, inspectable cues for telling informing, advising and consoling apart in prag's speech-act layer. - **I3** (learn): Mova could mine such per-label cue lists from silver data to propose candidate repair rules or explanations, then check them against held-out data. - **I4** (learn): when adding a component, Mova's evaluation could include a "component alone" run next to the full system, so that a weak combination step is not mistaken for useless knowledge. - **I5** (tree): when lexical or rule-based hints are injected into gparse, simple concatenation should not be assumed sufficient; ablations should check whether the hints are actually being used. *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 - [Northwestern University in Qatar](https://md.research.elamur.ai/free/en/universities/northwestern-university-in-qatar/) — University, Doha, Qatar · 24 works · 148 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: Mabrouka Bessghaier, Shimaa Ibrahim, Md. Rafiul Biswas and Wajdi Zaghouani. 2025. MarsadLab at AraHealthQA: Hybrid Contextual–Lexical Fusion with AraBERT for Question and Answer Categorization. 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{bessghaier2025marsadlab, title = {{MarsadLab at AraHealthQA: Hybrid Contextual–Lexical Fusion with AraBERT for Question and Answer Categorization}}, author = {Mabrouka Bessghaier, Shimaa Ibrahim, Md. Rafiul Biswas and Wajdi Zaghouani}, booktitle = {Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks}, year = {2025}, doi = {10.18653/v1/2025.arabicnlp-sharedtasks.33}, url = {https://doi.org/10.18653/v1/2025.arabicnlp-sharedtasks.33} } ``` - [← previous MarsadLab at AraGenEval Shared Task: LLM-Based Approaches to Arabic Authorship Style Transfer and Identification](https://md.research.elamur.ai/free/en/papers/biswas-2025-marsadlab-arageneval-shared-task-llm/) - [next → MarsadLab at BAREC Shared Task 2025: Strict-Track Readability Prediction with Specialized AraBERT Models on BAREC](https://md.research.elamur.ai/free/en/papers/ibrahim-2025-marsadlab-barec-shared-task-2025/) - [↑ back to papers of 2025](https://md.research.elamur.ai/free/en/papers/2025/page/15/) ## References 15 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. Arfan Ahmed, Nashva Ali, Mahmood Alzubaidi, Wajdi Zaghouani, Alaa Abd-alrazaq, and Mowafa Househ. 2022. Arabic chatbot technologies: A scoping review. Computer Methods and Programs in Biomedicine Update, 2:100057. [↗](https://doi.org/10.1016/j.cmpbup.2022.100057) 2. Firoj Alam, Shaden Shaar, Fahim Dalvi, Hassan Sajjad, Alex Nikolov, Hamdy Mubarak, Giovanni Da San Martino, Ahmed Abdelali, Nadir Durrani, Hassan Alhuzali, Ashwag Alasmari, and Hamad Al-saleh. 2024. Mentalqa: An annotated arabic corpus for questions and answers of mental healthcare. IEEE Access, 12:101155–101165. [find](https://scholar.google.com/scholar?q=Firoj+Alam%2C+Shaden+Shaar%2C+Fahim+Dalvi%2C+Hassan+Sajjad%2C+Alex+Nikolov%2C+Hamdy+Mubarak%2C+Giovanni+Da+San+Martino%2C+Ahmed+Abdelali%2C+Nadir+Durrani%2C+Hassan+Alhuzali%2C+Ashwag+Alasmari%2C+and+Hamad+Al%2Dsaleh%2E+2024%2E+M) 3. Hassan Alhuzali, Farah Shamout, Muhammad Abdul-Mageed, Chaimae Abouzahir, Mouath Abu-Daoud, Ashwag Alasmari, Walid Al-Eisawi, Renad Al-Monef, Ali Alqahtani, Lama Ayash, Nizar Habash, and Leen Kharouf. 2025. Arahealthqa 2025 shared task description paper. In Proceedings of ArabicNLP 2025. [find](https://scholar.google.com/scholar?q=Hassan+Alhuzali%2C+Farah+Shamout%2C+Muhammad+Abdul%2DMageed%2C+Chaimae+Abouzahir%2C+Mouath+Abu%2DDaoud%2C+Ashwag+Alasmari%2C+Walid+Al%2DEisawi%2C+Renad+Al%2DMonef%2C+Ali+Alqahtani%2C+Lama+Ayash%2C+Nizar+Habash%2C+and+Leen+Kharouf%2E) 4. N. Bensalah, H. Ayad, A. Adib, and A. I. El Farouk. 2024. Mindwave app: Leveraging ai for mental health support in english and arabic. In 2024 IEEE 12th International Symposium on Signal, Image, Video and Communications (ISIVC), pages 1–6. IEEE. [find](https://scholar.google.com/scholar?q=N%2E+Bensalah%2C+H%2E+Ayad%2C+A%2E+Adib%2C+and+A%2E+I%2E+El+Farouk%2E+2024%2E+Mindwave+app%3A+Leveraging+ai+for+mental+health+support+in+english+and+arabic%2E+In+2024+IEEE+12th+International+Symposium+on+Signal%2C+Image%2C+Video) 5. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. [Language models are few-shot learners](https://md.research.elamur.ai/free/en/papers/brown-2020-language-models-few-shot-learners/). Advances in neural information processing systems, 33:1877–1901. 6. Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. 2023. [Palm: Scaling language modeling with pathways](https://md.research.elamur.ai/free/en/papers/chowdhery-2022-palm-scaling-language-modeling-pathways/). Journal of Machine Learning Research, 24(240):1–113. [↗](https://doi.org/10.48550/arxiv.2204.02311) 7. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. [Bert: Pre-training of deep bidirectional transformers for language understanding](https://md.research.elamur.ai/free/en/papers/devlin-2019-bert-pre-training-deep-bidirectional/). In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186. [↗](https://doi.org/10.18653/v1/n19-1423) 8. Ridha Mezzi, Aymen Yahyaoui, Mohamed Wassim Krir, Wadii Boulila, and Anis Koubaa. 2022. Mental health intent recognition for arabic-speaking patients using the mini international neuropsychiatric interview (mini) and bert model. Sensors, 22(3):846. [↗](https://doi.org/10.3390/s22030846) 9. Ahd Mohamed and Wajdi Zaghouani. 2024. Expression of depression among arab twitter users using arabic corpus analysis. Procedia Computer Science, 244:76–85. 6th International Conference on AI in Computational Linguistics. [find](https://scholar.google.com/scholar?q=Ahd+Mohamed+and+Wajdi+Zaghouani%2E+2024%2E+Expression+of+depression+among+arab+twitter+users+using+arabic+corpus+analysis%2E+Procedia+Computer+Science%2C+244%3A76%E2%80%9385%2E+6th+International+Conference+on+AI+in+Compu) 10. Preslav Nakov, Alberto Barrón-Cedeño, Giovanni Da San Martino, Firoj Alam, Julia Maria Struß, Thomas Mandl, Rubén Míguez, Tommaso Caselli, Mucahid Kutlu, Wajdi Zaghouani, Chengkai Li, Shaden Shaar, Gautam Kishore Shahi, Hamdy Mubarak, Alex Nikolov, Nikolay Babulkov, Yavuz Selim Kartal, and Javier Beltrán. 2022. The clef-2022 checkthat! lab on fighting the covid-19 infodemic and fake news detection. In Advances in Information Retrieval, pages 416–428, Cham. Springer International Publishing. [find](https://scholar.google.com/scholar?q=Preslav+Nakov%2C+Alberto+Barr%C3%B3n%2DCede%C3%B1o%2C+Giovanni+Da+San+Martino%2C+Firoj+Alam%2C+Julia+Maria+Stru%C3%9F%2C+Thomas+Mandl%2C+Rub%C3%A9n+M%C3%ADguez%2C+Tommaso+Caselli%2C+Mucahid+Kutlu%2C+Wajdi+Zaghouani%2C+Chengkai+Li%2C+Shaden+Shaar%2C+Ga) 11. Hajar Sakai and Sarah S Lam. 2025. Large language models for healthcare text classification: A systematic review. arXiv preprint arXiv:2503.01159. [↗](https://arxiv.org/abs/2503.01159) 12. Chereen Shurafa and Wajdi Zaghouani. 2025. Corpus analysis of covid-19 related loneliness on twitter. In Arabic Language Processing: From Theory to Practice, pages 80–93, Cham. Springer Nature Switzerland. [find](https://scholar.google.com/scholar?q=Chereen+Shurafa+and+Wajdi+Zaghouani%2E+2025%2E+Corpus+analysis+of+covid%2D19+related+loneliness+on+twitter%2E+In+Arabic+Language+Processing%3A+From+Theory+to+Practice%2C+pages+80%E2%80%9393%2C+Cham%2E+Springer+Nature+Switzer) 13. Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, et al. 2022. [Lamda: Language models for dialog applications](https://md.research.elamur.ai/free/en/papers/thoppilan-2022-lamda-language-models-dialog-applications/). arXiv preprint arXiv:2201.08239. [↗](https://arxiv.org/abs/2201.08239) 14. Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. [Llama: Open and efficient foundation language models](https://md.research.elamur.ai/free/en/papers/touvron-2023-llama-open-efficient-foundation-language/). arXiv preprint arXiv:2302.13971. [↗](https://arxiv.org/abs/2302.13971) 15. N. Zahran, A. E. Fouda, R. J. Hanafy, and M. E. Fouda. 2025. [A comprehensive evaluation of large language models on mental illnesses in arabic context](https://md.research.elamur.ai/free/en/papers/zahran-2025-comprehensive-evaluation-large-language-models/). arXiv preprint arXiv:2501.06859. [↗](https://arxiv.org/abs/2501.06859) *Updated 1 November 2026*