Helping or Hurting? Predicting Changes in Users' Risk of Self-Harm Through Online Community Interactions

Luca Soldaini, Timothy Walsh, Arman Cohan, Julien Han, Nazli Goharian

Introduction

Online mental health communities are a double-edged sword. While they offer immediate, anonymous support for those in crisis, they also risk normalizing self-harm or spreading "contagious" suicidal ideation—a phenomenon known as the "Werther effect." Researchers from Georgetown University explored whether Natural Language Processing (NLP) can be used to predict the outcome of these digital interactions, specifically asking: Can we tell if a distressed user is being helped or harmed by the replies they receive in a forum thread?

Methodology and Dataset

The researchers framed the problem as a binary classification task. Given a thread that starts with a "flagged" (at-risk) post, the goal is to predict whether the user's final post in that thread will be "green" (not at-risk) or remain "flagged."

An example of a "flagged" post where a user expresses deep distress and suicidal ideation.

The study utilized the 2016 CLPsych ReachOut.com dataset, containing over 64,000 posts from an Australian mental health support forum. Due to a scarcity of manual annotations, a high-accuracy silver-standard labeling approach was used to create a dataset of 1,040 conversation threads.

A core technical innovation was Symmetric Modeling. Instead of analyzing the thread as a single block of text, the team extracted features for the target user (the one in distress) and the participating users (the responders) separately. Features included LIWC (Linguistic Inquiry and Word Count) for psychological states, TextBlob for sentiment, and LDA (Latent Dirichlet Allocation) for topic modeling.

Key Insights

The study revealed that modeling the distressed user and the responders separately significantly outperformed models that averaged their features. This suggests that the specific interaction between support types and distress types determines the outcome.

One surprising finding was the Engagement Paradox: a strong negative correlation ($\rho = -0.91$) between the number of times a target user posts and the likelihood of them reaching a "green" state. Longer threads often represented a "struggle to connect" rather than productive support.

A successful interaction where peer support leads a distressed user to a "green" state after calling a prevention hotline.

Furthermore, peer responses mentioning family, relationships, religion, and counseling were strongly associated with risk reduction. Conversely, when distressed users focused heavily on work or family-related distress, they were statistically less likely to overcome their state within that specific conversation.

Results and Performance

The Logistic Regression model achieved the highest macro-F1 score of 0.6910. While XGBoost provided higher precision (75.72), Logistic Regression offered a better balance for identifying both "helped" and "unhelped" users. LIWC features proved to be the most predictive, confirming that psychological linguistic markers are more telling than raw vocabulary in this context.

Conclusion

By quantifying these interactions, this research moves us closer to "smart" moderation tools. Such systems could flag when a conversation is failing to help, allowing human moderators to step in before a crisis escalates, ensuring online communities remain a safe space for peer support.

Summary updated

Abstract

In recent years, online communities have formed around suicide and self-harm prevention. While these communities offer support in moment of crisis, they can also normalize harmful behavior, discourage professional treatment, and instigate suicidal ideation. In this work, we focus on how interaction with others in such a community affects the mental state of users who are seeking support. We first build a dataset of conversation threads between users in a distressed state and community members offering support. We then show how to construct a classifier to predict whether distressed users are helped or harmed by the interactions in the thread, and we achieve a macro-F1 score of up to 0.69.
BibTeX
@inproceedings{soldaini2018helping,
  author = {Luca Soldaini and Timothy Walsh and Arman Cohan and Julien Han and Nazli Goharian},
  title = {{Helping or Hurting? Predicting Changes in Users' Risk of Self-Harm Through Online Community Interactions}},
  booktitle = {CLPsych @ NAACL},
  year = {2018},
  doi = {10.18653/v1/w18-0621},
}
Figure