DeTAILS：基于迭代式大语言模型支持的深度主题分析 (DeTAILS: Deep Thematic Analysis with Iterative LLM Support)

Thematic analysis is widely used in qualitative research but can be difficult to scale because of its iterative, interpretive demands. We introduce DeTAILS, a toolkit that integrates large language model (LLM) assistance into a workflow inspired by Braun and Clarke's thematic analysis framework. DeTAILS supports researchers in generating and refining codes, reviewing clusters, and synthesizing themes through interactive feedback loops designed to preserve analytic agency. We evaluated the system with 18 qualitative researchers analyzing Reddit data. Quantitative results showed strong alignment between LLM-supported outputs and participants' refinements, alongside reduced workload and high perceived usefulness. Qualitatively, participants reported that DeTAILS accelerated analysis, prompted reflexive engagement with AI outputs, and fostered trust through transparency and control. We contribute: (1) an interactive human-LLM workflow for large-scale qualitative analysis, (2) empirical evidence of its feasibility and researcher experience, and (3) design implications for trustworthy AI-assisted qualitative research.

翻译：主题分析在质性研究中被广泛使用，但由于其迭代性和解释性要求，难以规模化。我们介绍了DeTAILS，这是一个将大语言模型（LLM）辅助功能整合到受Braun和Clarke主题分析框架启发的工作流程中的工具包。DeTAILS通过旨在保持分析能动性的交互式反馈循环，支持研究人员生成和提炼编码、审查聚类以及综合主题。我们邀请了18位质性研究人员分析Reddit数据来评估该系统。定量结果显示，LLM支持的输出与参与者的提炼结果高度一致，同时减少了工作量并获得了较高的感知有用性。定性方面，参与者报告称DeTAILS加速了分析进程，促使他们对AI输出进行反思性参与，并通过透明度和控制机制建立了信任。我们的贡献包括：（1）一种用于大规模质性分析的交互式人-LLM工作流程；（2）其可行性及研究者体验的实证证据；（3）对可信赖的AI辅助质性研究的设计启示。