thematic coding

Qualitative thematic analysis (TA) faces scalability challenges due to its heavy reliance on researcher subjectivity and iterative reflexivity. This paper proposes a human-AI collaborative framework for AI-assisted TA, grounded in Braun & Clarke’s six-phase method. It leverages large language models (LLMs) to automate initial coding and preliminary theme generation, while an interactive interface enables researchers to provide real-time feedback and iteratively refine clustering, theme naming, and thematic synthesis. The framework preserves researcher agency and methodological transparency while enhancing analytical efficiency and reflective depth. An empirical user evaluation (N=18) demonstrates high alignment between AI-generated outputs and expert revisions, a 42% reduction in workload, a 3.1× speedup in analysis time, and a mean subjective usefulness rating of 4.6/5.0—validating its feasibility and practical utility for qualitative research.

Current open-source text analysis tools exhibit significant limitations in scalability, statistical modeling capabilities, and alignment with social science research paradigms, hindering paradigm-driven qualitative analysis of large-scale textual data. This study introduces an open-source Python framework designed specifically for computational social science, integrating sociological and anthropological research logic with scalable NLP architectures. It supports visual exploration and pattern discovery across heterogeneous qualitative sources—including field notes and web-based texts. Methodologically, the framework adopts a problem-oriented—rather than technology-driven—design, embedding core qualitative analysis workflows; leverages PyTorch and SciPy ecosystems to enable efficient distributed processing of document collections exceeding one million items; and provides low-code parameter interfaces alongside modular architecture to facilitate iterative, theory-informed validation bridging qualitative reasoning and quantitative modeling. The framework addresses critical gaps in the open-source ecosystem concerning openness, reproducibility, and methodological sensitivity.

Current large language models (LLMs) lack reproducible integration methods and systematic quality evaluation frameworks for qualitative software engineering research—particularly thematic analysis (TA). Method: We propose the first LLM-oriented prompt engineering and evaluation framework for TA, grounded in Braun & Clarke’s reflexive TA approach. It includes reproducible prompt templates and a blinded evaluation rubric aligned with established trustworthiness criteria (e.g., Lincoln & Guba). We conduct empirical evaluation across three models (GPT-4, Claude, Llama3) on 15 software engineer well-being interview transcripts. Contribution/Results: Expert blind assessment reveals that 61% of LLM-generated codes surpass human-coded ones in quality; however, critical limitations—including data fragmentation and meaning omission—are identified. The study establishes, for the first time, empirically informed boundaries for human–AI collaboration in qualitative analysis, providing both a methodological foundation and practical guidelines for LLM-augmented qualitative research.

Existing data-analytic models, grounded in positivism, neglect critical dimensions of power, tacit knowledge, and cognitive schemata. Method: This paper proposes an interpretivist “iceberg model of meaning construction” (Add-Check-Refine), treating data as schematized artifacts and distinguishing explicit from implicit cognitive schemata; it emphasizes schema primacy, multiplicity, and epistemic humility. Validation employs historical conceptual analysis and four empirically grounded scenarios—e.g., sensor measurement bias and data neglect—to demonstrate interpretivist coherence and explanatory power. Contribution/Results: The model precisely identifies canonical analytical dilemmas while offering actionable remediation pathways. Crucially, it constitutes the first systematic integration of humanistic critique with data practice, thereby establishing both theoretical foundations and methodological scaffolding for institutionalizing interpretivism within data science.

Qualitative data analysis in software engineering faces challenges including time intensity, poor reproducibility, and difficulty ensuring inter-rater reliability; the potential of large language models (LLMs) for human–AI collaboration in such tasks remains underexplored. This paper introduces the first explainable multi-agent framework tailored for qualitative research, enabling automated coding, theme extraction, and cross-textual synthesis via role-based task decomposition, prompt engineering, iterative validation, and a closed-loop human feedback mechanism. The architecture preserves human oversight and ensures analytical traceability, overcoming LLM limitations in low-shot, high-reliability settings. Empirical evaluation demonstrates a 3.2× improvement in analysis efficiency, scalability to hundreds of interviews, 89.7% accuracy in theme identification, and strong endorsement by domain experts.

This study addresses the inconsistent and poorly governed application of large language models (LLMs) in qualitative research, stemming from a lack of clear distinction between levels of meaning-making and modeling commitments. To resolve this, the paper proposes a 4×4 landscape framework that systematically integrates four meaning-making levels—description, categorization, explanation, and theorization—with four modeling commitments: static structure, sequential staging, causal pathways, and feedback dynamics. Through systematic literature mapping and multidimensional categorization, the analysis reveals that current LLM applications predominantly occupy lower-order meaning-making and low-commitment modeling, with scant engagement in explanatory or theoretical inference and dynamic mechanisms. This work establishes the first unified framework to delineate LLM capabilities and gaps in qualitative data analysis, offering a novel agenda and methodological foundation for governable AI-assisted qualitative research.

Chronic disease patient experience research relies heavily on qualitative thematic analysis, yet conventional approaches are labor-intensive, highly subjective, and difficult to scale. To address these limitations, we propose CoTI—the first collaborative multi-agent large language model framework designed specifically for qualitative analysis. CoTI comprises three specialized agents—Instructor, Thematizer, and CodebookGenerator—that jointly perform automated theme extraction, dynamic codebook generation, and alignment with expert judgment. Leveraging role-based prompt engineering, semantic similarity evaluation, and integrated human-in-the-loop interfaces, CoTI significantly enhances analytical efficiency and inter-rater consistency. Empirical evaluation on heart failure patient interviews demonstrates that CoTI’s thematic outputs achieve significantly higher agreement with senior researchers than those produced by junior researchers or baseline NLP models. The framework has been implemented as a deployable, production-ready analytical tool.

This study explores the effective integration of large language models (LLMs) into qualitative and mixed-methods social network analysis to augment—rather than replace—the deep analytical capacities of human researchers. Focusing on core issues such as relational meaning, narrative interpretation, and identity construction, the work proposes an LLM application paradigm oriented toward enhancing methodological rigor. This paradigm emphasizes human–AI collaboration, reflexive practice, and ethical accountability. By incorporating LLM-assisted data coding, theory generation, and abductive reasoning, the project develops a methodologically innovative yet practically feasible framework for qualitative social network analysis. The approach significantly improves analytical efficiency while upholding scholarly standards and ethical compliance.

This study investigates how generative artificial intelligence can be effectively integrated into qualitative research while preserving the researcher’s interpretive authority and analytical primacy. To this end, the authors propose HACITA, a human-AI collaborative framework, and develop a dedicated tool, ITA-GPT, which employs structured prompts to support familiarization with textual data, verbatim coding, descriptive coding using gerunds and nouns, and theme generation. The tool incorporates features for text tracing, coverage verification, and auditability to ensure transparency and traceability throughout the analytical process. Researchers retain continuous interpretive control through functionalities enabling modification, deletion, insertion, and annotation of AI-generated outputs. The study demonstrates the feasibility of organically blending automated assistance with human judgment in inductive thematic analysis, offering a novel paradigm for responsible human-AI collaboration in qualitative inquiry.

This study addresses the responsible integration of large language models (LLMs) into qualitative research workflows while upholding core epistemological principles such as reflexivity, situatedness, and interpretive judgment. By bridging qualitative methodology with explainable AI through an interdisciplinary lens, it systematically aligns key LLM technical parameters—including context window, temperature, top-p sampling, prompt design, and system cards—with the epistemological foundations of qualitative inquiry. The work demonstrates how LLMs differ fundamentally from traditional NLP tools in terms of transparency and interpretability. It further proposes a practical framework for qualitative researchers that explicitly links technical configurations to research ethics and methodological rigor, thereby advancing a critical and responsible synthesis of AI technologies with humanities and social science methodologies.