Priyanka Kargupta, Shuhaib Mehri, Dilek Hakkani-Tur, Jiawei Han

Sparking Scientific Creativity via LLM-Driven Interdisciplinary Inspiration

Identify insights from other domains that help address challenges or open up novel opportunities for your research problem:

1. Decomposes a research problem into questions.
2. Searches target-domain literature.
3. Searches cross-domain literature for transferable ideas.
4. Integrates and ranks cross-domain inspirations.

Quick Start (via inspiration_pred.py)

1) Install dependencies

pip install -r requirements.txt

2) Configure API keys

search.py reads Semantic Scholar credentials from config.py (API_KEY).

3) Run the pipeline

Default run (uses data/cross-domain-inspiration-relations.json):

python inspiration_pred.py

Useful options:

python inspiration_pred.py \
  --problem_file data/cross-domain-inspiration-relations.json \
  --model_name Qwen/Qwen3-14B \
  --output_dir inspiration_pred_output \
  --max_papers_per_query 20 \
  --temp 0.7 \
  --min_rel_threshold 0.5 \
  --skip_if_exists

4) Outputs

Results are written to inspiration_pred_output/*.json (or your custom --output_dir). Each output file contains:

• Problem metadata (research_problem, target_domain, fine_grained_domain, source_groundtruth)
• Cross-domain evidence grouped by question/domain
• idea_rankings (ranked integrated ideas)

5) Prepare evaluation inputs from ground-truth abstracts

If you want to convert ground-truth arXiv abstracts into this repo's evaluation format, use:

python evaluation/process_abstracts.py

Data Format: data/cross-domain-inspiration-relations.json

This file is a JSON array. Each entry is one cross-domain inspiration relation used as an input problem. The dataset is derived from: CHIMERA

Core fields used by inspiration_pred.py:

• source_id (int): source paper identifier
• target_id (int): target paper identifier
• source_domain (str): source domain (used as target/focus domain in this pipeline)
• target_domain (str): referenced inspired domain
• source_text (str): source-side idea phrase
• target_text (str): target-side inspiration phrase
• context (str): problem statement passed to decomposition
• publication_year (int): used to bound literature search
• abstract (str): stored as ground truth metadata

Additional metadata fields in the dataset:

• id, relation, arxiv_categories
• fine_grained_source_domain, fine_grained_target_domain
• head_leakage, tail_leakage
• paper_id

Minimal schema:

[
  {
    "id": "...",
    "source_id": 18243,
    "target_id": 38965,
    "source_domain": "Philosophy",
    "target_domain": "Computer Science",
    "source_text": "...",
    "target_text": "...",
    "relation": "inspiration",
    "publication_year": 2021,
    "paper_id": 2105.00867,
    "abstract": "...",
    "context": "..."
  }
]