<p>Early detection of depression and suicide risk from textual data is a critical challenge in mental health informatics, where false negatives can have severe real-world consequences. Transformer-based language models such as SentenceTransformer models based on MPNet have shown strong performance in affective text classification, yet they largely operate on isolated textual signals and overlook the broader contextual and relational patterns through which mental health distress often manifests. Graph-based representations offer a natural way to capture such relationships, but prior studies suggest that naive fusion of text and graph features frequently leads to limited or unstable improvements. In this work, we introduce a Context-Gated Graph Fusion framework that combines MPNet-based semantic sentence representations with era-aware graph centrality features through a learnable multiplicative gating mechanism. Rather than treating textual and relational information as equally informative, the proposed gate allows the model to rely on graph context only when it is supported by semantic evidence in the text. This design reflects a clinically intuitive assumption: contextual risk factors are meaningful only when grounded in linguistic expression. We evaluate the proposed approach on a dataset of 6,037 social media posts annotated with fine-grained emotional states and suicide-risk labels. The model achieves a Suicide Sensitivity (Recall) of 77.56%, outperforming standard text–graph concatenation baselines (66.67%) while maintaining stable F1 score, as well as transformer–GNN hybrid models and LLM-based baselines in zero-shot and few-shot settings. Cross-dataset evaluation on the Kaggle Suicide Watch dataset shows that the model generalizes well to unseen data. Additionally, ablation experiments (both encoder ablation and component-wise ablation) show that these gains arise from the gating mechanism itself, rather than the inclusion of graph features alone. For ensuring model robustness, stratified 5-fold cross-validation (average recall of 77.06%), McNemar’s statistical significance test (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(p < 0.05\)</EquationSource> </InlineEquation>), and a temporal “Time Capsule” analysis (revealing a 16.93% amplification of suicide risk in the post-pandemic era) were conducted. Together, these results highlight the value of context-aware fusion for safety-critical mental health prediction and support the practical relevance of the proposed framework.</p>

Temporal-semantics: an era-aware Context-Gated framework for suicide risk detection on social media

期刊论文
发表日期：

2026年6月17日

查看全文

Sneha,
Shreya Sisodia,
Tanisha Bansal,
Shatakshi Bansal,
Kiran Malik

摘要

Early detection of depression and suicide risk from textual data is a critical challenge in mental health informatics, where false negatives can have severe real-world consequences. Transformer-based language models such as SentenceTransformer models based on MPNet have shown strong performance in affective text classification, yet they largely operate on isolated textual signals and overlook the broader contextual and relational patterns through which mental health distress often manifests. Graph-based representations offer a natural way to capture such relationships, but prior studies suggest that naive fusion of text and graph features frequently leads to limited or unstable improvements. In this work, we introduce a Context-Gated Graph Fusion framework that combines MPNet-based semantic sentence representations with era-aware graph centrality features through a learnable multiplicative gating mechanism. Rather than treating textual and relational information as equally informative, the proposed gate allows the model to rely on graph context only when it is supported by semantic evidence in the text. This design reflects a clinically intuitive assumption: contextual risk factors are meaningful only when grounded in linguistic expression. We evaluate the proposed approach on a dataset of 6,037 social media posts annotated with fine-grained emotional states and suicide-risk labels. The model achieves a Suicide Sensitivity (Recall) of 77.56%, outperforming standard text–graph concatenation baselines (66.67%) while maintaining stable F1 score, as well as transformer–GNN hybrid models and LLM-based baselines in zero-shot and few-shot settings. Cross-dataset evaluation on the Kaggle Suicide Watch dataset shows that the model generalizes well to unseen data. Additionally, ablation experiments (both encoder ablation and component-wise ablation) show that these gains arise from the gating mechanism itself, rather than the inclusion of graph features alone. For ensuring model robustness, stratified 5-fold cross-validation (average recall of 77.06%), McNemar’s statistical significance test ( \(p < 0.05\) ), and a temporal “Time Capsule” analysis (revealing a 16.93% amplification of suicide risk in the post-pandemic era) were conducted. Together, these results highlight the value of context-aware fusion for safety-critical mental health prediction and support the practical relevance of the proposed framework.