Elamur Research⌕
☰
home / papers / 2026 / Assessing Suicide Risk in Arabic Crisis Helpline Calls: A Comparison of Arabic and English Large Language Models

Assessing Suicide Risk in Arabic Crisis Helpline Calls: A Comparison of Arabic and English Large Language Models

Summary

The paper asks whether suicide risk can be estimated automatically from Arabic crisis-helpline conversations without sensitive audio leaving the service. Working with Lebanon's National Lifeline (Embrace), the authors kept all recordings on site. There, a Whisper-family speech recognizer tuned for Levantine Arabic produced transcripts, and an Arabic named-entity recognizer removed identifying details. Only the redacted text reached the researchers. Operator-recorded C-SSRS ideation items were collapsed into two yes/no targets, at-risk and high-risk. Transcripts were also machine-translated into English, so the same calls with the same splits could be modelled in both languages. Five instruction-tuned LLMs were fine-tuned on each corpus with a single shared QLoRA recipe, and six transformer encoders served as baselines. High-risk detection worked better than the milder at-risk category. The best English model slightly outperformed the best Arabic one. This shows that a privacy-respecting, in-service pipeline for non-English crisis calls is workable. It also suggests that translate-then-classify need not cost accuracy.

Key points

  • The dataset has 383 eligible calls, each 5–10 minutes long. 373 had a determinable at-risk label (52.3% positive) and 297 had a determinable high-risk label (30.0% positive).
  • At-risk combines the two lowest C-SSRS rungs: a desire to die and active suicidal thoughts that lack specifics. High-risk combines the three rungs involving a method, intent, or plan.
  • On high-risk, the best Arabic model reached macro-F1 81.19 and ROC-AUC 90.61. The best English model reached 85.00 and 92.59 and caught 88.9% of high-risk calls.
  • Translation used Qwen/Qwen2.5-72B-Instruct with greedy decoding. 29 of 438 transcripts produced broken output and were removed from both corpora.
  • All models shared one 4-bit QLoRA setup (r = 16, α = 32, dropout 0.05) with no per-model hyperparameter search. The task was framed as an instruction whose expected answer is a single word.

Interesting ideas

  • I1. Keep a label only when it is determined. A label built by OR-ing several sub-items is kept only if the recorded items already settle it (any recorded item present, or every item recorded as absent). Otherwise the example is dropped rather than guessed as negative, which keeps systematic false negatives out of the data. — For Mova (learn): When silver labels are aggregated from partial witness-parser outputs, accept an aggregate only if the missing votes could not flip it, instead of treating absent votes as "no".
  • I2. Filter both sides of a paired comparison. If a transformation fails on some items, remove those items from every condition, not just the affected one. That way, measured differences come from the condition and not from a changed sample. — For Mova (back): In text → graph → text or translation experiments, evaluate original and round-tripped versions only on items where every pipeline variant succeeded.
  • I3. One fixed training recipe for all backbones. Applying identical hyperparameters to every model means score gaps reflect the backbone rather than tuning effort. — For Mova (tree): Compare the perceptron parser and gparse variants under a frozen training protocol, so treebank or architecture effects are not confused with tuning luck.
  • I4. Inspect the positive class, not just averages. Moderate macro-F1 or accuracy can hide degenerate models that label everything positive or nothing positive. Reporting positive-class recall and PR-AUC exposes this. — For Mova (learn): Add a check to evaluation that flags near-constant predictors, used as a negative control alongside aggregate scores for rare categories such as irony or specific speech acts.
  • I5. Collapse an ordered scale into severity groups. Grouping the rungs of an ordered scale into coarser bands gives tractable binary tasks while keeping their order. The study found the middle band harder to separate than the extreme one. — For Mova (frames): Graded pragmatic phenomena, such as the strength of an implicature or the force of a speech act, could be grouped into ordered bands, with extra attention paid to the subtle lower bands.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Linguistics & languages

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2609.00191 · arXiv 2609.00191

@misc{ma2026assessing,
  title = {{Assessing Suicide Risk in Arabic Crisis Helpline Calls: A Comparison of Arabic and English Large Language Models}},
  author = {Ma, Linhai and Hachem, Rita El and Hajj, Mahatab El and Ghandour, Lilian and Fodeh, Samah},
  year = {2026},
  doi = {10.48550/arxiv.2609.00191},
  eprint = {2609.00191},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2609.00191}
}

References

29 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. World Health Organization. Suicide worldwide in 2021: global health estimates. Geneva: World Health Organization; 2025. ISBN 978-92-4-011006-9. find
  2. Gould MS, Kalafat J, Harrismunfakh JL, Kleinman M. An evaluation of crisis hotline outcomes. Part 2: Suicidal callers. Suicide and Life-Threatening Behavior. 2007;37(3):338–352. https://doi.org/10.1521/suli.2007.37.3.338. ↗
  3. Zeinoun PA, Yehia FE, Khederlarian LZ, Yordi SF, Atoui MM, El Chammay R, et al. Evaluation of Lebanon’s National Helpline for Emotional Support and Suicide Prevention: reduction of emotional distress among callers. Intervention. 2021;19(2):197–207. https://doi.org/10.4103/INTV.INTV 50 20. ↗
  4. Posner K, Brown GK, Stanley B, Brent DA, Yershova KV, Oquendo MA, et al. The Columbia Suicide Severity Rating Scale: initial validity and internal consistency findings from three multisite studies with adolescents and adults. American Journal of Psychiatry. 2011;168(12):1266–1277. https://doi.org/10.1176/appi. ajp.2011.10111704. ↗
  5. Swedish Council on Health Technology Assessment. Instruments for Suicide Risk Assessment. Stockholm: Swedish Council on Health Technology Assessment (SBU); 2015. SBU Yellow Report No. 242. SBU Systematic Review Summaries. find
  6. Arowosegbe A, Oyelade T. Application of natural language processing (NLP) in detecting and preventing suicide ideation: a systematic review. International Journal of Environmental Research and Public Health. 2023;20(2):1514. https://doi.org/10.3390/ijerph20021514. ↗
  7. Castillo-Sánchez G, Marques G, Dorronzoro E, Rivera-Romero O, Franco-Martı́n M, De la Torre-Dı́ez I. Suicide risk assessment using machine learning and social networks: a scoping review. Journal of Medical Systems. 2020;44(12):205. https://doi.org/10.1007/s10916-020-01669-5. ↗
  8. Fodeh S, Goulet J, Brandt C, Hamada AT. Leveraging Twitter to better identify suicide risk. In: Proceedings of the First Workshop on Medical Informatics and Healthcare, 23rd ACM SIGKDD Conference on Knowledge Discovery and Data Mining. vol. 69 of Proceedings of Machine Learning Research; 2017. p. 1–7. Available from: https://proceedings.mlr.press/v69/fodeh17a.html. ↗
  9. Broadbent M, Medina Grespan M, Axford K, Zhang X, Srikumar V, Kious B, et al. A machine learning approach to identifying suicide risk among text-based crisis counseling encounters. Frontiers in Psychiatry. 2023;14:1110527. https://doi.org/10.3389/fpsyt.2023.1110527. ↗
  10. Iyer R, Nedeljkovic M, Meyer D. Using voice biomarkers to classify suicide risk in adult telehealth callers: retrospective observational study. JMIR Mental Health. 2022;9(8):e39807. https://doi.org/10.2196/39807. ↗
  11. Su Z, Jiang H, Yang Y, Hou X, Su Y, Yang L. Acoustic features for identifying suicide risk in crisis hotline callers: machine learning approach. Journal of Medical Internet Research. 2025;27:e67772. https://doi.org/10.2196/67772. ↗
  12. Wang H, Li J, Zhao Q, Chen Z, Song C, Tang J, et al. Deep learning-based feature fusion for emotion analysis and suicide risk differentiation in Chinese psychological support hotlines. arXiv preprint. 2025;arXiv:2501.08696. ↗
  13. Grimland M, Liberman M, Yeshayahu H, Benatov J, Munz N, Segal A, et al. Explainable AI for suicide risk detection: gender- and age-specific patterns from real-time crisis chats. Frontiers in Medicine. 2025;12:1703755. https://doi.org/ 10.3389/fmed.2025.1703755. ↗
  14. Maalouf FT, Alrojolah L, Akoury-Dirani L, Barakat M, Brent D, Elbejjani M, et al. Psychopathology in children and adolescents in Lebanon study (PALS): a national household survey. Social Psychiatry and Psychiatric Epidemiology. 2022;57(4):761–774. https://doi.org/10.1007/s00127-021-02208-4. ↗
  15. Baroud E, Ghandour LA, Alrojolah L, Zeinoun P, Maalouf FT. Suicidality among Lebanese adolescents: prevalence, predictors and service utilization. Psychiatry Research. 2019;275:338–344. https://doi.org/10.1016/j.psychres.2019.03.033. ↗
  16. Bizri M, Zeinoun L, Mihailescu AM, Daher M, Atoui M, Chammay R, et al. A closer look at patterns and characteristics of suicide in Lebanon: a first nationwide report of cases from 2008 to 2018. Asian Journal of Psychiatry. 2021;59:102635. https://doi.org/10.1016/j.ajp.2021.102635. ↗
  17. Farran D, El Haj M, Zarzour M, Chamoun Y, Kazazian B, Posbic P, et al. Factors associated with suicidal ideation and behaviour: analysis from the national suicide prevention helpline in Lebanon. BMC Psychiatry. 2025;25. https://doi.org/10. 1186/s12888-025-07602-9. ↗
  18. Radford A, Kim JW, Xu T, Brockman G, McLeavey C, Sutskever I. Robust speech recognition via large-scale weak supervision. In: Proceedings of the 40th International Conference on Machine Learning. vol. 202 of Proceedings of Machine Learning Research; 2023. p. 28492–28518. find our card
  19. Hu EJ, Shen Y, Wallis P, Allen-Zhu Z, Li Y, Wang S, et al. LoRA: low-rank adaptation of large language models. In: International Conference on Learning Representations (ICLR); 2022. . find
  20. Dettmers T, Pagnoni A, Holtzman A, Zettlemoyer L. QLoRA: efficient finetuning of quantized LLMs. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 36; 2023. . find
  21. Pappagari R, Zelasko P, Villalba J, Carmiel Y, Dehak N. Hierarchical transformers for long document classification. In: 2019 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). IEEE; 2019. p. 838–844. find
  22. Qwen Team. Qwen2.5 technical report. arXiv preprint. 2024;arXiv:2412.15115. ↗
  23. Grattafiori A, Dubey A, Jauhri A, Pandey A, et al. The Llama 3 herd of models. arXiv preprint. 2024;arXiv:2407.21783. ↗
  24. Huang H, Yu F, Zhu J, Sun X, Cheng H, Song D, et al. AceGPT: localizing large language models in Arabic. arXiv preprint. 2023;arXiv:2309.12053. ↗ our card
  25. Zhu J, Huang H, Lin Z, Liang J, Tang P, Yang X, et al. Second language (Arabic) acquisition of LLMs via progressive vocabulary expansion. In: Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers); 2025. ArXiv:2412.12310. ↗
  26. Inoue G, Alhafni B, Baimukan N, Bouamor H, Habash N. The interplay of variant, size, and task type in Arabic pre-trained language models. In: Proceedings of the Sixth Arabic Natural Language Processing Workshop (WANLP); 2021. p. 92–104. find our card
  27. Antoun W, Baly F, Hajj H. AraBERT: transformer-based model for Arabic language understanding. In: Proceedings of the 4th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT4), LREC 2020; 2020. p. 9–15. find our card
  28. Abdul-Mageed M, Elmadany A, Nagoudi EMB. ARBERT & MARBERT: deep bidirectional transformers for Arabic. In: Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics; 2021. p. 7088–7105. find our card
  29. Devlin J, Chang MW, Lee K, Toutanova K. BERT: pre-training of deep bidirectional transformers for language understanding. In: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics; 2019. p. 4171–4186. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/ma-2026-assessing-suicide-risk-arabic-crisis/