Mental health professionals and clinicians have observed the upsurge of mental disorders due to Interpersonal Risk Factors (IRFs). To simulate the human-in-the-loop triaging scenario for early detection of mental health disorders, we recognized textual indications to ascertain these IRFs : Thwarted Belongingness (TBe) and Perceived Burdensomeness (PBu) within personal narratives. In light of this, we use N-shot learning with GPT-3 model on the IRF dataset, and underscored the importance of fine-tuning GPT-3 model to incorporate the context-specific sensitivity and the interconnectedness of textual cues that represent both IRFs. In this paper, we introduce an Interpretable Prompting (InterPrompt)} method to boost the attention mechanism by fine-tuning the GPT-3 model. This allows a more sophisticated level of language modification by adjusting the pre-trained weights. Our model learns to detect usual patterns and underlying connections across both the IRFs, which leads to better system-level explainability and trustworthiness. The results of our research demonstrate that all four variants of GPT-3 model, when fine-tuned with InterPrompt, perform considerably better as compared to the baseline methods, both in terms of classification and explanation generation.

翻译：摘要：心理健康专业人士和临床医生观察到，因人际风险因素导致的心理障碍呈上升趋势。为模拟早期检测心理健康障碍的人工参与分诊场景，我们识别了个人叙事中反映两种人际风险因素的文本线索：受挫归属感与感知累赘感。基于此，我们在人际风险因素数据集上采用GPT-3模型的N-shot学习，并强调了微调GPT-3模型以纳入情境特异性敏感度及两类风险因素文本线索关联性的重要性。本文提出一种可解释提示方法——InterPrompt，通过微调GPT-3模型增强注意力机制。该方法通过调整预训练权重实现更精细化的语言修正，使模型能够学习识别两类人际风险因素的常见模式与潜在关联，从而提升系统层面的可解释性与可信度。研究结果表明，经InterPrompt微调后，GPT-3模型的全部四种变体在分类与解释生成任务中均显著优于基线方法。