The Role of Features and Context on Suicide Ideation Detection

The Role of Features and Context on Suicide Ideation Detection Yufei Wang∗ School of ITEE The University of Queensland Brisbane, Australia Yufei.Wang1@uq.net.au Stephen Wan and C´ecile Paris CSIRO Data61 Sydney, Australia firstname.lastname@data61.csiro.au Abstract There is a growing body of work study- ing suicide ideation, expressions of in- tentions to kill oneself, on social media. We explore the problem of detecting such ideation on Twitter, focusing on the impact of a set of features drawn from the liter- ature and on the role of discussion con- text for this task. Our experiments show a significant improvement upon the pre- viously published results for the O’Dea et al. (2015) dataset on suicide ideation. Interestingly, we found that stylistic fea- tures helped while social media metadata features did not. Furthermore, discussion context was useful. To further understand the contributions of these different fea- tures and of discussion context, we present a discussion of our experiments in varying the feature representations, and examining their effects on suicide ideation detection on Twitter. 1 Introduction According to World Health Organisation, a suicide occurs every 40 seconds worldwide (WHO, 2014). Suicidal death has destructive effect on both fam- ily (Cerel et al., 2008) and community (Levine, 2008) level. Tragically, many suicide cases can be prevented (Bailey et al., 2011). As social me- dia platforms, such as Twitter1, are often used as channels to discuss mental health topics, there is a need for new technologies to deliver online men- tal health support (Daine et al., 2013). Such ser- vices may be particular important for the youth, well represented on social media, for whom sui- cide is the second leading cause of death (WHO, 2014). ∗This work was performed while Yufei Wang was on a CSIRO Student Vacation Scholarship, 2015-2016 1www.twitter.com Consequently, there is a growing body of work that studies suicide ideation, expressions of inten- tions to kill oneself, on platforms such as Twit- ter. For example, O’Dea et al. (2015) describe a data set of Twitter posts that has been annotated by mental health and social media experts for (i) the presence of suicide ideation, and (ii) the level of severity of the ideation. In that text classifica- tion work, lexical features alone were used. How- ever, intuitively, one might expect that informa- tion, such as the discussion context, might each provide valuable information to detect cases of suicide ideation. For example, information from the surrounding discussion context, perhaps by friends, might indi- cate the presence of genuine suicide ideation. Two examples, Post A and Post B and their respective replies, are shown below.2 Post-A: Okay goodbye, im going to kill myself tomorrow @ the retreat thing. Reply-A: @ANON No plz dont. Post-B: Listening to ultra live stream rn in ANON’s car da gonna kill myself Reply-B: @ANON I was watching it at work!! Although both cases contain the key phrase “kill myself”, the replies indicate that Post-A is a more concerning post than Post-B, as the respondent answers sympathetically and supportively. How- ever, the reply to …

Content truncated to 500 words