Cost Sensitive Psychological Crisis Warning Analysis Based on Big Language Model Dual System Cognition and Feature Space Reshaping

Abstract

In the context of psychological counseling, timely and accurate identification of visitors' suicide and self injury risks is the core of crisis intervention. However, existing research mostly focuses on static single sentence classification and lacks perception of the dynamic evolution trajectory of emotions; Meanwhile, due to the extremely low proportion of high-risk samples in real data, general large language models commonly suffer from security overload after being aligned through reinforcement learning based on human feedback. Faced with extreme distress signals, they are prone to trigger exemption and refusal to answer, resulting in fatal false negatives. In response to the above challenges, this study proposes a cost sensitive early warning architecture based on ChatGLM3-6B that integrates dual system cognition and feature space reshaping. From the aspects of architecture design, fine-tuning strategy, loss function design, etc., it provides an effective and ethically feasible technical paradigm for constructing a zero false alarm and high-precision intelligent psychological intervention system.