Detecting Suicide Risk in Online Counseling Services: A Study in a Low-Resource Language

Proceedings of the 29th International Conference on Computational Linguistics, pages 4241–4250 October 12–17, 2022. 4241 Detecting Suicide Risk in Online Counseling Services: A Study in a Low-Resource Language Amir Bialer† Daniel Izmaylov† Avi Segal† Oren Tsur† {amirbial,zmaylov}@post.bgu.ac.il avisegal@gmail.com orentsur@bgu.ac.il Yossi Levi-Belz‡ yossil@ruppin.ac.il Kobi Gal†§ kobig@bgu.ac.il †Ben-Gurion University of the Negev ‡Ruppin Academic Center §University of Edinburgh Abstract With the increased awareness of situations of mental crisis and their societal impact, online services providing emergency support are be- coming commonplace in many countries. Com- putational models, trained on discussions be- tween help-seekers and providers, can support suicide prevention by identifying at-risk indi- viduals. However, the lack of domain-specific models, especially in low-resource languages, poses a significant challenge for the automatic detection of suicide risk. We propose a model that combines pre-trained language models (PLM) with a fixed set of manually crafted (and clinically approved) set of suicidal cues, followed by a two-stage fine-tuning process. Our model achieves 0.91 ROC-AUC and an F2-score of 0.55, significantly outperforming an array of strong baselines even early on in the conversation, which is critical for real-time detection in the field. Moreover, the model performs well across genders and age groups. 1 Introduction The World Health Organization (WHO) lists sui- cide as one of the most salient causes of death world-wide, causing more deaths than breast can- cer or war (WHO, 2019). With close to 1 million lives taken directly by suicide every year, and over 25 million suicide attempts, suicide incurs a lasting impact on families and communities. Identifying individuals at risk ahead of time and providing them with psychological and medical support is a key step in suicide prevention (Joiner et al., 2018). In this paper, we tackle the task of detecting Sui- cide Ideation (SI). Specifically, we aim at the detec- tion of SI of individuals contacting online counsel- ing hot-lines in low-resource and morphologically- rich languages. We focus on anonymous data from online, text-based (chat), support services. Such services are available in many countries, allowing for confidential and immediate help to those in distress, and play a critical role in suicide preven- tion (Bantilan et al., 2021; Jashinsky et al., 2014; Joiner et al., 2007). Empirical evidence suggests that at-risk individuals seek help in close proximity to actual suicide attempts (Zalsman et al., 2021), thus it is critical to identify suicide risks as early as possible during the session. There is a growing body of work on assessing suicide risk from English texts (whether in social media posts or from counseling sessions), but there is an acute lack of NLP resources that could be used for detection of suicide risk in other languages (Lee et al., 2020). We directly address this gap by focusing on suicide risk detection from online coun- seling services in Hebrew, which is a low-resource and morphologically-rich language that challenges traditional NLP tasks (Seker et al., 2021). Our proposed approach for the SI detection task, called SI-BERT, extends a generic pretrained model with a small set of Out …

Content truncated to 500 words