Sentiment aware prompts improve mental health chatbot responses
PromptSense: Sentiment-Aware Mental Health Chatbot with Dynamic Prompting
Artificial Intelligence
Summary
Mental health chatbots often miss emotional cues and safety concerns, which can lead to less helpful or even harmful conversations. This paper shows a way to make chatbots better by first analyzing the user's emotions and depression signs, then adjusting the chatbot's prompts accordingly. The authors tested different methods and found that their approach improves how empathetic, clear, and safe the chatbot's replies are. This could help make mental health chatbots more supportive and trustworthy.
What this means in practice
- •For mental health app developers: Create chatbots that respond more empathetically and safely by dynamically adjusting prompts based on user emotions and risk signals.$Commercial implications: Enables development of improved mental health chatbot products with better user engagement and safety, appealing to healthcare providers and app markets.
- •For customer support teams: Integrate emotion-aware prompting in support chatbots to improve response relevance and reduce unsafe or toxic outputs.
Authors
Sugi Lee, Jianlong Zhou, Jing sUN
Abstract
Current Large Language Models (LLMs)-based mental health chatbots often lack sufficient empathy, contextual awareness, and safety mechanisms, potentially leading to ineffective or harmful interactions. This paper proposes a novel framework that integrates sentiment and depression analysis with a rule-based filtering mechanism prior to model inference. By dynamically constructing prompts based on emotional and risk-related insights, the framework aims to significantly improve the contextual relevance, empathy, and risk-aware of model-generated responses. We evaluated four model strategies and two prompt engineered variants across qualitative metrics such as empathy, helpfulness, clarity, relevance, safety, as well as toxicity assessments using the Perspective API. Experimental results indicate that the prompt engineering approaches outperforms DSFT and RLAIF by 1.8\% and 1.1\% respectively.