ilog3 Education AI Assisted Qualitative Analysis
Goal
Use AI to accelerate qualitative analysis while preserving human interpretation, traceability, and research rigor.
Use After
Use with or after:
education-qualitative-coding-analysiseducation-qualitative-study-designeducation-mixed-methods-designeducation-action-research-designeducation-design-based-researcheducation-program-evaluationeducation-policy-comparative-analysis
Do not expose the skill name to users. Present it as "AI 辅助质性分析" or "AI 辅助编码建议".
Inputs
- Research questions
- Anonymized qualitative corpus or excerpts
- Existing codebook, if any
- Chosen analysis approach: thematic analysis, grounded theory, content analysis, case analysis
- Whether multiple coders will review AI suggestions
- Privacy constraints and whether cloud AI tools are allowed
Core Principle
AI can suggest, cluster, retrieve, and challenge. It cannot decide final codes, themes, or findings.
All AI outputs must be marked as:
- suggested
- accepted
- revised
- rejected
Workflow
- Privacy check:
- confirm text is anonymized
- remove names, school identifiers, student numbers, phone numbers, account IDs
- avoid cloud tools if data cannot leave local environment
- Prepare analysis batch:
- 3-10 excerpts for pilot coding
- or full corpus split into document/segment IDs
- Generate initial code suggestions:
- descriptive codes
- in vivo codes
- process codes
- values/emotion codes if relevant
- Ask researcher to accept/revise/reject codes.
- Cluster accepted/revised codes into categories.
- Generate theme candidates with definitions and evidence links.
- Search for negative cases:
- excerpts that contradict the dominant theme
- participants/cases that differ
- missing voices or edge cases
- Generate analytic memos:
- what pattern appears
- what evidence supports it
- what remains uncertain
- what needs human review
- Produce audit trail:
- prompt/task
- input batch IDs
- AI suggestion
- human decision
- revision note
- Update codebook and evidence table.
Tool Calls
Qualitative Software With AI Features
ATLAS.ti: https://atlasti.com/
MAXQDA AI Assist: https://www.maxqda.com/
NVivo: https://lumivero.com/products/nvivo/
Dedoose: https://www.dedoose.com/
Local / Controlled Analysis
Use local scripts or private LLM infrastructure when sensitive education data cannot be sent to web tools.
Suggested local preprocessing:
rg -n "姓名|学校|电话|身份证|学号|家长|微信|邮箱" transcripts/
Python text clustering support:
pip install pandas scikit-learn sentence-transformers umap-learn hdbscan
Use clustering only to suggest groupings; do not treat clusters as final themes.
Prompt Templates
Initial Coding Suggestion
你是教育研究中的质性分析助手。请根据研究问题，对以下匿名访谈片段提出初始编码建议。
要求：
1. 每个代码必须对应具体片段。
2. 区分描述性代码、过程代码、情感/价值代码。
3. 不要生成最终主题。
4. 标注不确定之处。
研究问题：
[RQ]
片段：
[excerpt_id + text]
Theme Clustering
请将以下已人工确认或修订的代码聚合为候选类别和候选主题。
要求：
1. 保留每个代码的来源。
2. 说明聚合理由。
3. 标注边界模糊的代码。
4. 不要删除少数/反常代码。
代码表：
[code table]
Negative Case Search
以下是一个候选主题。请在材料中寻找可能削弱、反驳、限制或复杂化该主题的片段。
候选主题：
[theme]
材料：
[excerpt list]
输出：
反例片段、为什么构成反例、是否需要修改主题。
Output Format
1. AI Suggested Codes
| Excerpt ID | Suggested Code | Code Type | Evidence Text Summary | Confidence | Human Decision | Revision |
|---|
Human decision values:
- Accept
- Revise
- Reject
- Needs discussion
2. Candidate Theme Clusters
| Candidate Theme | Included Codes | Rationale | Supporting Excerpts | Boundary/Negative Cases | Human Decision |
|---|
3. Negative Case Table
| Theme | Negative/Boundary Excerpt ID | Why It Matters | Suggested Theme Revision |
|---|
4. Audit Trail
| Date | Task | Input IDs | Tool/Model | Output Type | Human Reviewer | Decision Summary |
|---|
5. Memo Template
Memo ID:
Related RQ:
Related codes/themes:
Pattern:
Evidence:
Negative cases:
Interpretive risk:
Next analytic action:
Quality Rules
- Never call AI-suggested codes "final codes".
- Never report a theme without evidence excerpts.
- Always preserve rejected AI suggestions in the audit trail when the decision affects analysis.
- AI may help find negative cases; the researcher decides whether they alter the theme.
- For minor students, anonymize text more aggressively because writing samples can be identifiable.
- For policy/document analysis, distinguish text frequency from policy importance.
- For classroom discourse, preserve interaction sequence before clustering.
User-Facing Closure
End by asking the user to choose a review mode, not a tool:
我可以先做两种方式之一：A. 对 3-5 份材料做试编码，帮助你建立初始代码本；B. 基于已有代码本做主题聚类和反例搜索。你想先走哪一步？