arXiv cs.AIThursday · May 28, 2026FREE

SuiChat-CN: Benchmarking Contextual Suicide Risk Assessment in Chinese Group Chats

SuiChat-CN addresses the gap in suicide risk assessment for instant messaging environments like Telegram, where messages are short, fragmented, and culturally specific. The benchmark was constructed by collecting public Telegram group-chat data, extracting coherent conversational segments via signal-word extraction and bidirectional context expansion, and annotating user risk levels with an expert-validated, LLM-assisted paradigm. The dataset contains 13,312 contextual segments from 1,406 users, covering 258,228 raw chat messages. Extensive experiments with PLMs and over 40 LLMs demonstrate that contextual information is essential for reliable risk assessment, while fine-tuning and partial-context evaluation significantly improve performance over isolated post-level analysis. This work highlights the need for context-aware models in suicide prevention on group chat platforms.