Suicidal Risk Detection for Military Personnel

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 2523–2531, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 2523 Suicidal Risk Detection for Military Personnel Sungjoon Park* 1, 2, Kiwoong Park* 1, Jaimeen Ahn 1 Alice Oh 1 1 School of Computing, KAIST, Republic of Korea 2 Upstage AI Research, Upstage, Republic of Korea {sungjoon.park, marspak, jaimeen01}@kaist.ac.kr alice.oh@kaist.edu Abstract We analyze social media for detecting the sui- cidal risk of military personnel, which is es- pecially crucial for countries with compulsory military service such as the Republic of Korea. From a widely-used Korean social Q&A site, we collect posts containing military-relevant content written by active-duty military per- sonnel. We then annotate the posts with two groups of experts: military experts and men- tal health experts. Our dataset includes 2,791 posts with 13,955 corresponding expert anno- tations of suicidal risk levels, and this dataset is available to researchers who consent to re- search ethics agreement. Using various fine- tuned state-of-the-art language models, we pre- dict the level of suicide risk, reaching .88 F1 score for classifying the risks. 1 Introduction Suicide is one of the major causes of death in the military. In some countries where military service is compulsory because of a conscription system, active-duty military personnel live in physical sepa- ration from their family and friends for an extended period of time, often against their will. In the Re- public of Korea, for example, most men have the obligation to serve in the military for about a year and half, leading to a large population of about 600,000 in active duty as of this year. Many of them experience difficulty adapting to the isolated environment, and some of them are at risk of sui- cide. One approach to detect the suicide risk signs of active-duty soldiers is the analysis of social media posts, similar to the approach used for detecting sui- cide risk of the general public (Milne et al., 2016; Yates et al., 2017; Zirikly et al., 2019). However, * These authors contributed equally. research on military suicide finds that there are dis- tinct risk factors, such as combat exposure, injury, bereavement, and negative unit climate associated only with military service (Nock et al., 2013). For this reason, we cannot directly apply the findings of suicide risk research of the general public to the military personnel. In this paper, we take on the challenge of collecting social media posts related to military service in the Republic of Korea, anno- tating and analyzing them using NLP methods for detecting suicide risk of active-duty soldiers. The first and most challenging step is to cre- ate an annotated dataset of military-related social media posts written by active-duty soldiers. We collect posts from a popular social question and answering (Q&A) platform. Anonymous posts are allowed, so we find that there is a considerable amount of military-related posts that contain pos- sible suicide risk and other mental health issues. Annotation poses a challenge, as the mental health and suicide …

Content truncated to 500 words