Phrase-level pairwise topic modeling to uncover helpful peer responses to online suicidal crises

Abstract

Abstract Suicide is a serious public health problem; however, suicides are preventable with timely, evidence-based interventions. Social media platforms have been serving users who are experiencing real-time suicidal crises with hopes of receiving peer support. To better understand the helpfulness of peer support occurring online, this study characterizes the content of both a user’s post and corresponding peer comments occurring on a social media platform and present an empirical example for comparison. It introduces an approach that uses pairwise topic models to transform large corpora of discussion into associated topics of user and peer posts. The key advantages of this approach include: (i) modeling both the generative process of each type of the corpora (i.e., user posts and peer comments) and the associations between them, and (ii) using phrases, which are more informative and less ambiguous than words, in addition to words, to represent social media posts and topics. The study evaluated the method using data from Reddit r/SuicideWatch. It examined how the topics of user and peer posts were associated and how this information influenced the peer perceived helpfulness of the peer support. Then, this study applied structural topic modeling to data collected from individuals with a history of suicidal crisis as a means to validate findings. The observations suggest that effective modeling of the association between the two lines of topics can uncover helpful peer responses to online suicidal crises, notably providing the suggestion of professional help. The proposed technology can be applied to “paired” corpora in many applications, such as technical support forums, question-answering sites, and online medical services.