Abstract
Early detection of suicidal ideation on social media is a critical component of modern suicide prevention. However, many languages, such as Korean, lack the expert-validated "suicide dictionaries" required for high-performance detection models. This paper presents a novel cross-lingual framework that leverages clinical knowledge from high-resource languages (English and Chinese) to identify suicidal risk in Korean social media posts. By employing specialized Suicidal-Oriented Word Embeddings (SOWE) and an ensemble architecture, the proposed model achieves 87.50% accuracy, demonstrating that cross-lingual transfer can effectively bridge the gap in mental health resources.
Introduction
Social media platforms have emerged as vital frontiers for identifying individuals at risk of suicide. Effective detection models typically rely on domain-specific knowledge, often in the form of "suicide dictionaries"—curated lists of high-risk terms validated by mental health experts. While such resources exist for high-resource languages like English and Chinese, they are notably absent for many others, including Korean. This resource gap hinders the development of robust suicide prevention tools in these linguistic contexts.
Proposed Framework
To address this, we propose a cross-lingual ensemble model designed to detect suicidal ideation in Korean social media posts by leveraging existing clinical knowledge from English and Chinese. The pipeline involves four primary stages:
- Translation: Korean posts are translated into English and Chinese using machine translation (Naver Papago).
- Suicidal-Oriented Word Embedding (SOWE): Standard embeddings are refined to distinguish between suicidal and non-suicidal expressions. For English and Chinese, this refinement is guided by expert dictionaries. For Korean, a computationally generated dictionary based on TF-IDF scores is utilized.
- Post Attention Layer: Each language-specific representation is processed through an LSTM with an attention mechanism to highlight risk-relevant sequences.
- Ensemble Layer: The final representations are concatenated, allowing the model to capture a diverse spectrum of risk signals across the three semantic spaces.
Data and Experimental Setup
The target dataset comprises approximately 31,000 posts from "Naver Cafe," a major Korean social media platform, with 10,000 labeled as suicide-related. Expert dictionaries for Chinese (2,168 words) and English (7,286 sentences from Reddit) provided the necessary domain knowledge for embedding refinement. For the Korean component, where no expert dictionary was available, the authors successfully employed a statistical approach using TF-IDF to identify high-risk keywords.
Results and Analysis
The ensemble model, CL-ours, achieved an accuracy of 87.50% and an F1-score of 87.49%, significantly outperforming the monolingual Korean baseline (85.74%). Analysis of the top keywords across languages revealed a mix of universal and culturally specific markers.
As shown in Figure 4, only 27 words were common to all three languages (e.g., "die," "family," "sad"). Korean-specific keywords often reflected local social pressures, such as "grades" and "beauty," whereas the ensemble approach allowed the model to integrate these local nuances with broader clinical signals from the English and Chinese models. This suggests that while suicidal ideation has a universal core, its linguistic expression is deeply tied to cultural context.
Conclusion
This study demonstrates that cross-lingual transfer is a viable strategy for suicide detection in resource-constrained languages. By treating translation as a feature and employing an ensemble of specialized embeddings, we can effectively utilize global clinical knowledge to improve local intervention efforts. Future work will focus on expanding the framework to additional languages and refining the computational dictionary generation process to further reduce reliance on manual expert curation.
Summary updated