【作者】 董王昊； 黄英辉； 周瑾宜； 王伟军；

【Author】 Wanghao Dong;Yinghui Huang;Jinyi Zhou;Weijun Wang;School of Psychology, Central China Normal University;School of Management, Wuhan University of Technology;

【机构】 华中师范大学心理学院； 武汉理工大学管理学院；

【摘要】 情感支持(Emotional Support)旨在减少情感困扰,帮助求助者理解并克服所面临的挑战。人工智能能否为人类提供情感支持,是长久以来备受关注的话题。ChatGPT是当前最具影响力的生成式人工智能,其突出的语言输出能力和前后文结合能力使得AI情感支持成为可能。本研究以ChatGPT为对象,通过设定专门的提示策略,发掘与验证其在情感支持中的能力。研究一:聚焦心理健康问答情景,目的在于探讨ChatGPT在单轮对话中的情感支持能力。基于由10903提问和19682回答组成的数据集,构建心理健康问答感知有用性的预测模型;通过特征敏感性分析及解释性机器学习方法,建立感知有用性的解释性模型,以识别影响感知有用性的关键性性语言线索及其影响方式,进而制定提示策略;然后采用统计分析方法比较提示之后的ChatGPT与人类咨询师在心理健康问答中的感知有用性。结果显示:在严重风险和中度风险问题上,提示后的ChatGPT达到了"人类咨询师"的平均水平(p> 0.05);而对于低风险和无风险问题,提示后的ChatGPT已经超过了"人类咨询师"的平均水平(6.80%和4.63%, p <0.001)。研究二:聚焦基于文本的线上咨询情景,目的在于探讨ChatGPT在多轮对话中的情感支持能力。以一个公开发表的,经过人工标注的,包含1300个案例的情感支持会话数据集为研究材料。利用基于深度学习的自然语言处理模型(BERT-Integrated Attention LSTM Model,BIALM),建立会话策略预测模型和情绪强度预测模型。前者可对ChatGPT进行针对性会话策略提示,后者可用以评估ChatGPT会话的有效性。然后对提示后的ChatGPT和人类专家的情感会话有效性进行分析比较。结果发现:经过会话策略提示的ChatGPT的表现显著高于人类咨询师的平均水平(13.25%, p <0.001)。最后,还对ChatGPT在不同问题类型上的表现进行了对比与探讨,并指明了存在的不足以及未来展望。总体上,本研究以计算社会科学为导向,初步证实了ChatGPT的情感支持能力,为AI在心理健康领域的运用提供了有力支持。更多还原

【Abstract】 Emotional support aims to alleviate emotional distress and assist individuals in understanding and overcoming the challenges they face. Whether artificial intelligence can provide emotional support to humans has been a long-standing and widely debated topic. ChatGPT, the most influential generative AI model, possesses remarkable language output and context comprehension capabilities, making AI emotional support feasible. The current research focuses on ChatGPT and employs specially designed prompt strategies to explore and validate its ability to provide emotional support. Study 1: Focusing on the context of mental health Q&A, aims to explore the emotional support capabilities of ChatGPT in single-turn dialogues. A predictive model for perceived usefulness is constructed based on datasets comprising 10,903 questions and 19,682 answers. Feature sensitivity analysis and explainable machine learning(XML) methods are employed to establish an interpretable model for perceived usefulness. This model identifies key linguistic cues and their impact on perceived usefulness, enabling the formulation of prompt strategies. Afterward, statistical analysis was employed to compare the perceived usefulness of prompted ChatGPT with that of human counselors. The results indicate that for high-and moderate-risk questions, prompted ChatGPT reaches the average level of "human counselors"(p > 0.05). However, for low-and no-risk questions, prompted ChatGPT even surpasses the average level of "human counselors"(6.80% and 4.63%,p < 0.001) in terms of perceived usefulness. Study 2: Focusing on the context of text-based online counseling, aims to explore the emotional support capabilities of ChatGPT in multi-turn dialogues. The research material consists of publicly available, manually annotated emotional support conversation datasets containing 1300 cases. Using a deep learning-based natural language processing model(BERT-Integrated Attention LSTM Model, BIALM), we establish two predictive models: a session strategy prediction model and an emotion intensity prediction model.The former is utilized to provide targeted session strategy prompts to ChatGPT, while the latter is employed to evaluate the effectiveness of ChatGPT’s sessions. Subsequently, an analysis and comparison are performed on the emotional conversation effectiveness of prompted ChatGPT and human counselors. The results indicated that prompted ChatGPT’s performance significantly surpasses the average level of human counselors(13.25%, p <0.001) regarding emotional conversation effectiveness. Finally, the performance of ChatGPT on different question types is also compared and discussed, and deficiencies as well as future perspectives are indicated. Overall, driven by computational social science, our study has provided preliminary evidence of ChatGPT’s emotional support capabilities, offering substantial support for the application of AI in the field of mental health.更多还原