Keyword Co-Occurrence Analysis

Author keywords are higher quality but sparse (not all papers have them). Text extraction is comprehensive but noisier. Best practice: use both. Author keywords as primary (quality), then supplement with extracted keywords for papers without assigned keywords. Always standardize and filter extracted keywords to remove stop words and noise. Report which approach was used.

Create a synonym dictionary manually for major keywords. For 'machine learning', group: ['machine learning', 'ML', 'statistical learning', 'supervised learning' (if context is ML not formal statistics)]. For 'neural networks', group: ['neural networks', 'deep learning', 'artificial neural networks']. Be conservative: 'neural networks' and 'artificial intelligence' overlap but are distinct—keep them separate unless your analysis specifically requires merging. Use domain expertise; do not automate standardization without expert review.

It depends on corpus size. For a corpus of 1000 papers, a threshold of 3–5 co-occurrences is typical. For 10,000 papers, 10–20 is reasonable. For 100,000+ papers, 50+ may be needed. The goal is to reduce noise while retaining signal. A rule of thumb: use a threshold that yields 100–500 keyword pairs; fewer pairs suggests your threshold is too high, more suggests too low. Start conservative (threshold=5) and relax if needed based on network density.

Not directly. Keyword co-occurrence shows what researchers are currently studying, not what will become important. However, temporal analysis (keyword growth trends) can suggest momentum: rapidly rising co-occurrence suggests growing community interest. Combine with citation analysis: rising keywords + rising citations = strong indicator of emerging prominence. Keywords alone cannot predict influence—many trending keywords prove to be temporary fads.