27 March 2017
Emotions.Tech are using machine learning algorithms to rank, list and search web pages according to emotion.
It's called the Emotional Web. Most popular is the "love" page: emotions.tech/discover/love, a daily listing of the happiest news from across the internet. Clicking on an article shows an emotional breakdown of the text.
Emotional analysis is also built into their Angry Blocker browser extension. This plugin for Google Chrome prevents the user from seeing angry or sad web pages, instead showing a picture from nature and a link back to the Emotional Web.
"We're trying to clean up the internet." says co-founder Paul Tero. "We think we can programmatically prevent emotional abuse and hate speech."
They are also interested in emotional targeting, a technology which can prevent ads appearing on inappropriate pages, something Google wish they had used in hindsight after the recent uproar about major brands advertising on extremist web sites.
Behind the scenes is a multi-layer recurrent neural network. When analysing a piece of text, Emotions.Tech's software first turns the words into high dimensional vectors (numbers basically) and then feeds it through their network. Millions of calculations later, the output is 5 numbers representing the amount of love, laughter, surprise, sadness and anger in the text. Running on an NVIDIA Titan Pascal graphics card, we can process each paragraph in less than a millisecond, about 50 times faster than in CPU. This approach is very different from most other companies providing sentiment or emotion analysis, who generally count the frequency of happy or sad words and calculate a score from that.
Emotions.Tech's analysis is also done from the reader's point of view, not the writer. This is a subtle but significant distinction. Browsing the Emotional Web's laughter page reveals many stories which classic sentiment analysis would probably score as negative, but which readers thought were hilarious - useful information for headline and ad writers.
Emotions.Tech are only a few months old, but have lofty ambitions. "We are deep learning's answer to propaganda" explains CEO Ilia Zaitsev.