Keyword Extraction for arXiv

Guidelines for turning fuzzy research ideas into effective arXiv search keywords.

Principles

1. Technical over generic: Prefer domain nouns (e.g. "transformer", "reinforcement learning") over vague terms ("research", "study").
2. 3–5 keywords: Enough to narrow scope; avoid single overly broad or many overlapping terms.
3. No stop words: Exclude "the", "a", "and", "for", "with" — arXiv handles these.
4. Combine concepts: If idea spans multiple areas, include 1–2 terms per area (e.g. "federated learning" + "privacy" + "healthcare").
5. Use arXiv categories when inferable: cs.AI, cs.LG, stat.ML for ML; cs.CL for NLP; etc.

Common Patterns

Retry Triggers

If EvaluatePapers returns score < 3, consider:

• Narrowing keywords (add domain or constraint)
• Replacing generic terms with technical synonyms
• Splitting one broad idea into 2–3 focused keyword sets and searching separately

Examples

• "用深度学习做医学图像分析" → ["deep learning", "medical image", "diagnosis", "CNN"]
• "大模型在代码生成上的应用" → ["large language model", "code generation", "programming"]
• "联邦学习隐私保护" → ["federated learning", "privacy", "differential privacy"]