Word clouds

011 min

Why teams reach for a word cloud

Picture a product designer with 600 written answers to one survey question: what would you change about the app? Reading them all would take days. So the designer pastes the answers into a free generator and, a few seconds later, gets a picture. The biggest words are app, use and like. The designer puts the picture on a slide for the next review.

A product manager looks at it and asks what the team should change. The picture has no answer. It shows that people wrote app often. It does not show whether they love the app or cannot open it.

That gap, between a quick look and a finding you can act on, is the main thing to understand about word clouds. To see why it exists, start with where the form came from.

022 min

Where word clouds came from

Sizing words by importance is older than the web. A map does it when it prints a large city in big type and a village in small type. Designers call this a weighted list. A printed example from 1995 appears in Douglas Coupland's novel Microserfs, which carries a list of English keywords sized this way.

The web version arrived in 2004. The first tag clouds on a large website appeared on the photo-sharing site Flickr in 2004, built by its co-founder Stewart Butterfield. A tag is a short label that someone attaches to a photo, page or link, and a tag cloud shows those labels at a size that follows how many items carry each one. Butterfield based the design on Jim Flanagan's Search Referral Zeitgeist, a display of the sites that sent visitors to a page. Del.icio.us and Technorati spread the idea soon after.

For a few years tag clouds were a common feature of blogs and early social sites. Then the mood turned. Flickr accepted its 2006 Webby Award for best practices with a five-word speech: "sorry about the tag clouds."

How Wordle turned tag clouds into word clouds

The word cloud, a picture of the words in any text, started as a side project. In 2004 the programmer Jonathan Feinberg and a colleague, Bernard Kerr, built a bookmarking tool at work, and like every tool of that time it showed a tag cloud. Feinberg found tag clouds dull and saw little evidence that they helped people navigate.

Then a blogger, Matt Jones, posted his own bookmark tags as a lively, carefully typeset image. Feinberg wanted a program that could make such images. He dropped the tags, counted the words in any text, and built Wordle. He has written that he designed it for pleasure rather than analysis. By the time he wrote about it, people had saved more than 1,400,000 clouds in its gallery.

Jacob Harris, a senior software architect at The New York Times, later argued that Wordle's ease had a cost. It made word clouds so simple to produce that news sites began using them as filler pictures.

032 min

How a word cloud is built

Every generator follows the same steps, whatever its pictures look like.

1. Split the text into words.
2. Drop stop words, the very common words such as the, it and to. Left in, they would be the biggest words in almost every cloud.
3. Count how often each remaining word appears.
4. Turn each count into a font size.
5. Place the words on the page so that none overlap.

The count in step 3 is a choice. A frequency cloud sizes a word by how often it occurs. A significance cloud sizes a word by how unusual it is compared with a large body of background text, so that common words do not win. A categorization cloud sizes a label by how many items sit under it, as in a website's tag list. The same technique also draws numbers other than counts: a data cloud sizes names by a figure such as population. All of them need only two things, a list of words and one number for each.

Position therefore carries no meaning. Nothing about where app sits tells you anything about app. Older web tag clouds took the opposite approach. They set the tags in lines, usually in alphabetical order, so a reader could find a tag by its first letter, or they sorted the tags by weight. Some layouts place related words near each other.

042 min

What word size can tell a reader

Now the picture exists, and the question is what a reader can take from it. The answer is size, and size alone.

Readers see area, and area is hard to compare. Marti Hearst and Daniela Rosner studied tag clouds at the University of California, Berkeley. They found that it is difficult to compare tags of similar size. Gregor Aisch, then at The New York Times, adds that colour is just as hard to decode, and that position is spent on filling space.

052 min

What tests found about word clouds and lists

Do people understand a text better from a cloud than from a plain list? Few researchers have tested it, and the tests are small. Hearst and Rosner reviewed them in 2008.

Hearst and Rosner conclude that clouds are probably not the best tool for navigation or for understanding abstract information. They also call the experimental work limited.

062 min

What a word cloud hides

Return to the product manager's question. A cloud cannot answer it, and Jacob Harris shows why with a much larger data set.

In 2011 Harris set two pictures from the same data side by side. The data was the WikiLeaks Iraq War Logs, the field reports from the war. One picture was his team's map of civilian deaths in Baghdad. The other was a word cloud of report titles published by Fast Company. He argued that word clouds support only the crudest sorts of textual analysis.

Equal sizes for car and blast might mean many car bombs, or many reports about cars and many about explosions. Reports about one kind of attack might say truck, vehicle or bongo, and a cloud counts each as a different word. A cloud also gives no context: nothing tells a reader that LN means Local National or that COP means Combat Outpost. And it gives no narrative. The one thing a reader learns is that the war involved many IEDs and explosions, which Harris says is news to nobody.

The same limits apply to the survey. A cloud counts single words, so a complaint made of several words is split apart. Slow, laggy and takes ages count as three separate things.

072 min

When a word cloud is the right choice

None of this makes the word cloud useless. It makes it a tool for one job.

Aisch set out that job in 2015. A word cloud is fine, he wrote, if the aim is to get people to read a big group of words that have little to do with each other, as long as each is tied to one or two interesting values. A word count from a text does not count as an interesting value. His own case used Zillow data that gave a median home value for each combination of street name and street suffix. Readers could scan hundreds of street names in a cloud in a way they would not read a list of the same names.

There is a second reason people choose clouds, and it is not about measuring. Hearst and Rosner studied why designers use them. People like clouds because they seem personal and social. A cloud suggests what a person or group is interested in and how that changes over time. It looks lively, and it takes less room than a long list. A cloud therefore works as a portrait of interests more than as a measurement.

The same logic explains live polling. Tools such as Mentimeter show a cloud that grows as the audience submits words, which suits a warm-up or a brainstorm. The point is to see the room's mood, not to count exactly. On a website, a tag cloud can also link each tag to its pages, so it doubles as a way to browse. The tests above suggest it is not the fastest way.

082 min

Making a word cloud that does not mislead

If a cloud is the right choice, a few moves keep it honest.

1. 1

   Clean before counting.

   Merge plurals, capital letters and spellings of the same word, and count short phrases such as dark mode as one item. Harris says a phrase or thematic analysis reaches more accurate conclusions than single words.

2. 2

   Check the stop word list.

   Keep negations such as not and don't, because dropping them can reverse the meaning of the top words.

3. 3

   Say what was counted.

   A caption such as "600 survey answers, common words removed" tells readers what the sizes mean.

4. 4

   Put the numbers next to it.

   A bar chart of the ten most frequent words lets readers compare counts, which they cannot do with font size. For words that belong to groups, a treemap or circle packing shows each word inside its group.

5. 5

   Guide the reader.

   Aisch limited his words to two angles, split them into three groups, added filters and wrote a few sentences to lead readers through the graphic. He also suggests word bubbles, where circle size is easier to compare than font size and position can carry another value.

6. 6

   Group first, then count.

   For the survey, sorting the answers into themes with affinity diagramming and counting the themes answers the manager's question better. If the question is how people feel, sentiment analysis measures it directly.

Giving position a meaning

A newer technique called Word Rain keeps the cloud and gives position a job. It turns each word into a list of numbers that reflects how the word is used, called a word embedding, and reduces the list to one value. That value sets the word's place along the horizontal axis, so words with close meaning sit near each other. A word's height follows how prominent it is, and less prominent words drop down until they no longer overlap a more prominent one. The result is still one static picture, but a neighbour now means something.

091 min

Word clouds and accessibility

A word cloud usually reaches readers as an image, and an image passes on each word's size only to people who can see it. WCAG 2.2 success criterion 1.3.1, Info and Relationships, aims to keep what visual formatting conveys when the page is read by a screen reader. If size is the only place a word's weight is recorded, a screen reader user gets none of it. Give the counts as text, in a list or table beside the cloud, or build the cloud from real text with the weight in the markup. Web tag clouds have long been built from text rather than images for a related reason: it makes them readable by software.

Colour needs care too. Success criterion 1.4.3, Contrast (Minimum), asks that text and its background differ by a contrast ratio of 4.5:1 or more, or 3:1 for large text. The smallest words in a cloud are the likeliest to fail. Check them against the background, and see contrast for how the ratio works.

101 min

Word clouds vs. nearby charts

A word cloud is easy to confuse with a tag cloud, because the picture is the same. The difference is the data. A tag cloud shows labels that people attached to items, such as photos. A word cloud shows words found in a text. Many people use the two names for one thing.

The more useful comparison is with the charts that answer the same question better.

Size in a word cloud works as a visual hierarchy that the data sets for you, and size is one of the preattentive attributes the eye notices before it reads. That is why clouds look informative at a glance and say so little once you read them.

?4 questions

Questions people ask

§10 sources

Sources