A Text-Based Suicide Detection Model Using Hybrid Prompt Tuning in Few-Shot Scenarios

Suicide is more prevalent among individuals with psychiatric disorders, underscoring the importance of early identification of warning signs for intervention. Common suicide detection models for text analysis often require tremendous labeled data, making them prone to overfitting when dealing with tiny datasets. Aiming at the problem, we propose a prompt-based learning suicide detection model that is suitable in low-resource settings following the “pre-train, prompt, predict” paradigm, named E3.0-HP-SDM (ERNIE 3.0 Hybrid Prompt-Suicide Detection Model). In the construction of the E3.0-HP-SDM, we selected ERNIE 3.0, renowned for its knowledge enhancement capabilities, as our pre-trained language model (PLM). Additionally, we developed a hybrid prompt template, which integrates a set of tunable soft prompts into a specific suicide-related hard prompt template. This template reformulates the original input into a format with unfilled slots, specifically designed to guide the PLM in applying its knowledge-masked language model for the inference of suicide intentions. When tested on identical data, E3.0-HP-SDM outperforms not only other models within the same paradigm but also often-cited baseline combination models that follow the third paradigm of natural language processing, the “pre-train, fine-tune” paradigm, with an accuracy of 87.6% and an AUC of 85.2%.