Code for the systematic literature search underpinning the review paper "Imbuing Creativity into Human-AI Systems: A Neurosymbolic Approach" (submitted to Nature Machine Intelligence).
The search followed PRISMA guidelines and the PRISMA-ScR extension for scoping reviews. Two databases were queried across five thematic clusters: neurosymbolic AI & creativity, cognitive architecture & dual-process theory, uncertainty & exploration in LLMs, ensemble & compositional methods, and psychedelic neuroscience & creativity.
code/
├── google-scholar/
│ ├── 1-keyword-search/ # SerpAPI keyword search + 18 query CSVs
│ │ └── queries/ # One CSV per search query
│ ├── 2-merge-deduplicate/ # Merge query results and remove duplicates
│ ├── 3-automated-prescreening/ # Automated inclusion/exclusion criteria
│ └── 4-chatgpt-screening/ # ChatGPT third-rater screening (GPT-5.2)
│
└── pubmed/
├── 1-semantic-search/ # BigQuery SQL — semantic vector search on PubMed Central
├── 2-deduplicate/ # Deduplication by PMC ID
├── 3-automated-prescreening/ # Automated inclusion/exclusion criteria
├── 4-chatgpt-screening/ # ChatGPT third-rater screening (GPT-5.2)
└── 5-gemini-crossvalidation/ # Gemini cross-validation (gemini-2.0-flash)
| Step | Folder | Description |
|---|---|---|
| 1 | 1-keyword-search/ |
Fetches results for 18 keyword queries via SerpAPI |
| 2 | 2-merge-deduplicate/ |
Merges all query CSVs and removes duplicates |
| 3 | 3-automated-prescreening/ |
Applies automated inclusion/exclusion (year, keywords, venue, preprint check) |
| 4 | 4-chatgpt-screening/ |
Sends borderline records to ChatGPT for third-rater screening |
| Step | Folder | Description |
|---|---|---|
| 1 | 1-semantic-search/ |
BigQuery SQL using text-embedding-005 semantic search across 5 cluster queries |
| 2 | 2-deduplicate/ |
Cleans and deduplicates by PMC ID |
| 3 | 3-automated-prescreening/ |
Automated screening with semantic distance thresholds (include < 0.82, exclude > 0.88) |
| 4 | 4-chatgpt-screening/ |
ChatGPT screening of borderline records |
| 5 | 5-gemini-crossvalidation/ |
Gemini cross-validation for PubMed candidates |
| Source | Initial | After dedup | After auto-filter | Manual review | Included |
|---|---|---|---|---|---|
| Google Scholar | 1,240 | 1,113 | — | 33 | 24 |
| PubMed Central | 232 | 232 | — | 46 | 19 |
| Combined | 1,472 | 1,345 | 79 | 79 | 43 |
Manual screening used a 2-of-3 majority vote across three independent raters: two reviewers and one AI reviewer (ChatGPT, GPT-5.2; OpenAI). Disagreements were resolved by a designated tiebreaker reviewer. See the paper's Supplementary Methods for the full protocol and prompt template.
# Google Scholar pipeline
pip install serpapi pandas
# PubMed pipeline
pip install google-cloud-bigquery pandas
pip install openai # part-4
pip install google-generativeai # part-5
API keys required: SERPAPI_KEY, OPENAI_API_KEY, GOOGLE_API_KEY, and a configured BigQuery project with access to the PubMed Central dataset.
Include:
- Peer-reviewed journal articles or conference proceedings
- Published 2020–2025 (2015+ for foundational neuroscience)
- ≥ 3 keywords matching the five thematic clusters
- Top-tier or well-regarded academic venue
Exclude:
- Editorials, commentaries, preprints, grey literature, dissertations
- Semantic distance > 0.88 (PubMed auto-exclusion)
- < 2 thematic keywords
MIT License. See LICENSE for details.