Data Mining Reveals How Wording Influences Tweet Propagation

If you’ve ever painstakingly crafted a tweet in the hope it would be retweeted around the world, only to find it flopped, then read on…

Anybody with a Twitter account will have spent time crafting the perfect tweet, designed to resonate with followers and force them to retweet it around the world. These same tweeters will also be painfully aware that the most popular tweets can sometimes be the ones that have had the least attention and were retweeted for reasons that are difficult to fathom.

That raises an interesting question. Does all this attention to editorial detail make any difference? Would your tweets be just as successful (or unsuccessful) if you fired them off with different messages?

It’s easy to imagine that it’s more or less impossible to work out what might have been had you phrased your tweets differently. But it turns out that there is a way to test how differently phrased tweets pointing to the same content have fared.

A surprising feature of Twitter is that there are many pairs of tweets from the same author that have different wordings but point to the same URL. Today, Chenhao Tan at Cornell University in Ithaca and a few pals, use a huge number of these pairs of tweets to study the question of what kind of wording makes a tweet successful.

In the process, they’ve discovered a set of features that distinguish highly shared tweets from those that are less popular.

Tan and co begin by crawling through the tweets posted by 236,000 users looking for pairs. Somewhat surprisingly, this process yielded 1.8 million pairs of tweets, each sent by the same author on the same topic (ie pointing to the same URL) and differing in more than just word spacing.

Twitter prevents users from sending the same tweet twice in quick succession. But tweeters get around this by adding spaces to their messages. So Tan and co also found 630,000 pairs of tweets that were identical apart from spacing.

That turned out to be handy because it allowed them to check how various factors influence the number of retweets. These factors include the number of followers, the time between each pair of tweets and the amount of information in a tweet.

This data indicated that they ought to study only those tweets from authors with more than 5000 followers and those sent within 12 hours of each other. Having filtered the database for these kinds of tweets, they were left with 11,404 pairs along with the number of times each was retweeted.

One interesting question is whether humans can tell which of the pair received the most retweets. To answer this question, Tan and co used Amazon’s Mechanical Turk service to present people with random pairs of tweets and asked them to choose the more popular option, and to give their reasons.

In fact, they have set up a website where anyone can have a go at: http://chenhaot.com/retweetedmore/quiz.

They found that humans successfully pick the more popular phrasing with an average accuracy of 61.3 per cent. “Not that high, but better than chance, indicating that it is somewhat possible for humans to predict greater message spread from different deliveries of the same information,” say Tan and co.

But these guys have gone even further by analysing the structure and content of the tweets and designing an algorithm that can learn which tweets do best.

In particular, this algorithm searched for various phrases, positive or negative sentiment, requests to share and so on. And this algorithm achieved the success rate of about 66 per cent, somewhat better than humans.

So what is the secret of the perfect tweet? Tan and co say better tweets are ones that include more information clearly, use language that aligns with previous messages and with the community at which it is aimed and ones that mimic news headlines in their structure. Requesting followers to retweet also helps.

That’s not a magic formula by any means. But it is certainly useful. If diligently applied, these are the kind of tips that could help to improve many a tweeter’s tweets.

Ref: arxiv.org/abs/1405.1438 : The Effect Of Wording On Message Propagation: Topic- And Author-Controlled Natural Experiments On Twitter

Follow the Physics arXiv Blog on Twitter at @arxivblog, on Facebook and by hitting the Follow button below.