In most walks of life, claiming that you could see Twitter’s face would have you locked up. Thankfully there is one situation where it is acceptable: academic research.
Said visualization was the conclusion of a presentation I recently attended by Nello Cristianini, the Professor of Artificial Intelligence at the University of Bristol’s Department of Engineering Mathematics. He works mainly on automated analysis in two areas: genomic data and, as he discussed here, media content.
Cristianini’s team at the university’s Intelligent Systems Laboratory uses seven servers to analyze the online output of around a thousand news outlets each day, which covers a tenth of the estimated total. The system receives approximately 50 articles each minute, with the daily total being the equivalent of two hundred 350-page books.
Originally the project only looked at media outlets. One of the techniques used for analysis involves parsing the text to try to reduce each story to a simple subject, object and verb. Cristianini’s team have been using this to look at media reports about the US presidential race and produced the chart below.
Both the size of the circle and its vertical position shows how prominent a candidate was in the media at a particular time. The horizontal position shows whether the person is more frequently the subject (usually the person making comments about others) or the object. The color show whether the person is making positive (blue) or negative (red) comments.
The laboratory staff have now also turned their attention to media created by the public, specifically on Twitter postings. Cristianini noted that analyzing individual tweets is extremely difficult as they come out at a far more rapid pace than mainstream media, and they are far shorter. However, working with a huge number of tweets makes it possible to detect patterns.
To test this theory, the researchers gathered together 484 million tweets made by British users over the past three years. They then searched for words and phrases that indicated particular moods (known as sentiment analysis.)
The results showed distinct increases in positive moods at Christmas, New Year and Valentine’s Day, along with a smaller increase at the time of last year’s Royal Wedding. Though predictable, these patterns suggest the idea that the mood of the Twitter user base can be measured.
There were two other clear patterns that were surprising, if only in their intensity. Around October 2010, when the UK government announced a major program of public spending cuts (which would likely lead to large number of job losses), tweets with a gloomy tone became more prevalent. And last August, as several British cities were hit by rioting, angry tones came to the forefront.
The most interesting part of this was that many of these depressed or enraged tweets weren’t directly mentioning the specific events. Instead the mood could be detected in posts discussing other elements of the poster’s lives.
After first displaying this as a series of graphs showing each sentiment’s respective levels over time, the team created an even simpler way to show the patterns: aggregating the various moods into an animated face that changed expression accordingly.
And thus, the face of Twitter: