Word clouds that actually say something useful

Word clouds have a bad reputation among data people, and mostly they earned it. Used carefully they are still the fastest way to sense the shape of a corpus you have never read.

Remove stop-words first, always

Without filtering, every cloud is a monument to "the", "and" and "of". A standard English stop-word list removes roughly forty percent of tokens and leaves the content words behind.

Watch for near-duplicates

"Test", "tests" and "testing" get counted separately and each looks minor, when together they may be the dominant theme. Scan the frequency table, not only the picture, before drawing conclusions.

Area exaggerates difference

Font size scales linearly but perceived weight scales with area, so a word twice as frequent looks roughly four times as important. Treat the cloud as a ranking, and export the CSV when you need the real numbers.

Good uses

• Skimming survey free-text responses before coding them
• Comparing two versions of a document
• Finding filler words in your own writing
• Building a visual for a talk where precision is not the point