【作者】 黄英辉； 董王昊； 周瑾宜； 王伟军；

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

【机构】 武汉理工大学管理学院； 华中师范大学心理学院；

【摘要】 近年来网络心理健康问答社区发展迅速,为公众提供了便捷、普惠且具有高度隐私保护的心理健康信息及服务。针对这种基于文本的异步心理健康问答服务的有效性和影响机制的研究尚且不足。本研究构建了一个心理健康问答情境下咨询师回答有效性的评估模型。首先,我们构建了包含19682条来访者求助及咨询师应答信息的问答数据集,以公众对咨询师回复的感知有用性作为咨询师问答质量的评价指标,采用适用于大规模问答文本分析的计算语言学和解释性人工智能(XAI)方法,建立咨询师回答感知有用性的可解释性预测模型;分析了求助者、咨询师及两者互动同步性三种来源的语言线索要素对回答的感知有用性的影响,并探讨了语言线索要素之间的交互效应。预测模型的性能评估结果显示,所构建的感知有用性预测模型的RMSE及MAPE值分别达到0.8234、22.7288%。特征消融实验的结果表明,咨询师卷入度相关语言线索(回答的词数和响应时间)具有最大的预测力(R2=0.1114),且情感联结、治疗任务一致性、治疗目标一致性等数字化治疗联盟相关语言线索能有效提高预测模型的效果。基于SHAP值(Shapley可加性解释)的解释性分析结果表明,咨询师积极的情感表露、卷入度、治疗任务一致性及治疗目标一致性相关语言线索正向预测回答有效性。因果推断分析结果进一步显示,上述预测关系受到咨询师回答的词数和响应时间的显著影响。本研究利用了可解释性机器学习及因果推断方法,从精细可能性模型及数字化治疗联盟理论视角,解释并验证不同心理语言学线索在单次、异步网络心理健康问答效果的预测作用,发现了咨询师的卷入度及数字化治疗联盟相关语言线索对其回答质量的显著预测作用,验证了咨询师卷入度对数字化治疗联盟相关语言线索的异质性效应,为理解社会化问答情境下心理支持服务效果的影响机制,提高心理健康问答服务质量提供了理论和方法参考。更多还原

【Abstract】 The rapid development of online mental health Q&A communities provides the public with convenient,inclusive, and highly confidential mental health information and services. However, research on the effectiveness and mechanism of such text-based asynchronous mental health Q&A services is still inadequate. This study aims to construct a model to assess the effectiveness of counselors’ responses in mental health Q&A scenarios. We compiled a Q&A dataset of 19,682 instances of helpseekers’ inquiries and counselor responses, and used the perceived helpfulness of counselor responses as the evaluation for the quality of the counselor’s response. We employed computational linguistics, machine learning, and causal inference methods suitable for large-scale Q&A text analysis to establish an interpretable predictive model for the perceived usefulness of counselors’ responses.We analyzed the influences of help seekers, counselors, and the synchronicity of their interactions, as well as three types of language cues-Linguistic Inquiry and Word Count(LIWC), interaction synchronicity, and peripheral cues-on the perceived usefulness of the responses, and explored the interactive effects among these language cues. The performance of our predictive model showed that the Root Mean Square Error(RMSE) and Mean Absolute Percentage Error(MAPE) of the constructed perceived usefulness predictive model were 0.8234 and 22.7288%respectively. The results of the feature ablation indicated that the language cues related to counselor involvement had the greatest predictive power(R2=0.1114), and the language cues related to emotional connection, therapeutic task consistency, and therapeutic goal consistency associated with the digital therapeutic alliance could effectively improve the effectiveness of the predictive model. The analysis results based on Shapley Additive Explanations(SHAP) values further demonstrated a positive prediction of answer effectiveness by the language cues related to counselor involvement and emotional connection, therapeutic task consistency, and therapeutic goal consistency.Causal inference analysis results showed that the aforementioned predictive relationships are significantly influenced by the counselor’s word count and response time difference. This research leverages explainable artificial intelligence methods to interpret and validate the predictive roles of various psycholinguistic cues in single-instance, asynchronous online mental health Q&A, from the perspectives of elaboration likelihood model and digital therapeutic alliance theory. It has found the significant predictive roles of counselor involvement and digital therapeutic alliance-related language cues on the Q&A quality, verified the heterogeneous effects of counselor involvement on digital therapeutic alliance-related language cues, providing theoretical and methodological references for understanding the mechanisms of influence on the effectiveness of psychological support services in social Q&A scenarios and improving the quality of mental health Q&A services.更多还原