[Paper Review] Using Linguistic Features to Estimate Suicide Probability of Chinese Microblog Users

Lei Zhang, Xiaolei Huang+3 more|arXiv (Cornell University)|Nov 4, 2014

Mental Health via Writing18 references18 citations

TL;DR

This study proposes a machine learning approach to predict suicide probability among Chinese microblog users using linguistic features extracted from Sina Weibo posts. By applying LIWC and LDA-based topic modeling to 1,041 users' microblogs, the authors show that inferred topics outperform LIWC features and achieve a best RMSE of 11.68, with inferred topics demonstrating equal or superior predictive power to trained topics while being more reusable.

ABSTRACT

If people with high risk of suicide can be identified through social media like microblog, it is possible to implement an active intervention system to save their lives. Based on this motivation, the current study administered the Suicide Probability Scale(SPS) to 1041 weibo users at Sina Weibo, which is a leading microblog service provider in China. Two NLP (Natural Language Processing) methods, the Chinese edition of Linguistic Inquiry and Word Count (LIWC) lexicon and Latent Dirichlet Allocation (LDA), are used to extract linguistic features from the Sina Weibo data. We trained predicting models by machine learning algorithm based on these two types of features, to estimate suicide probability based on linguistic features. The experiment results indicate that LDA can find topics that relate to suicide probability, and improve the performance of prediction. Our study adds value in prediction of suicidal probability of social network users with their behaviors.

Motivation & Objective

• To develop a scalable, proactive suicide risk detection system using social media data.
• To investigate whether linguistic features from microblogs can predict suicidal ideation with high accuracy.
• To compare the predictive performance of LIWC-based features versus topic modeling using LDA.
• To evaluate whether topics inferred from high-risk user groups outperform those trained on full datasets.
• To explore the reusability and generalization potential of inferred topics in mental health risk prediction.

Proposed method

• Collected Sina Weibo posts and SPS scores from 1,041 users to form a labeled dataset.
• Applied the Chinese edition of LIWC to extract lexical features related to emotional and cognitive states.
• Used Latent Dirichlet Allocation (LDA) to extract topic features from microblog texts.
• Trained supervised machine learning models (stepwise linear regression) using LIWC features, trained topics, and inferred topics.
• Compared model performance using RMSE as the metric across different feature combinations and topic numbers.
• Evaluated both trained topics (from full dataset) and inferred topics (from high-risk subgroup) for predictive power and reusability.

Experimental results

Research questions

• RQ1 Can linguistic features from Chinese microblog posts predict suicide probability with acceptable accuracy?
• RQ2 Do topic models derived from LDA outperform traditional lexical analysis (LIWC) in predicting suicidal ideation?
• RQ3 Is there a significant difference in predictive performance between topics trained on the full dataset versus topics inferred from a high-risk subgroup?
• RQ4 Can inferred topics—trained on a small, high-risk subset—achieve comparable or better performance than trained topics?
• RQ5 Does combining LIWC and topic features improve prediction accuracy, and if so, under what conditions?

Key findings

• The best-performing model achieved an RMSE of 11.68 using inferred topics with 70 topics, significantly outperforming the LIWC baseline (RMSE = 15.43).
• LDA-based topic modeling improved prediction accuracy compared to LIWC, confirming that topic features capture more nuanced psychological signals than lexical counts.
• Inferred topics achieved equal or higher predictive power than trained topics, especially at higher topic counts, and showed greater stability.
• The maximum correlation between a single topic and SPS score reached 0.15, with no significant improvement at higher topic numbers, suggesting diminishing returns beyond 70–100 topics.
• Combining LIWC with inferred topics did not improve performance, while LIWC+trained topics showed marginal gains only at higher topic counts.
• Inferred topics were more reusable than trained topics, as they could be applied to new data without retraining on the full dataset.