What’s he building in there?
WHO KNOWS WHAT other people dream of? Unless you’re a character in Inception, your dreams are for you and you alone. That’s good for you if your dreams involve your best friend’s girlfriend, but bad for scientists who want to study dreams using the scientific method.
Since dreams are not directly observable, you have to collect thousands of dream reports and depend on the reliability of the subjects’ recollection in order to systematically study their content.
To make matters worse, each of the dream reports must be analyzed. To do this, an expert judge must evaluate the content and code the accounts into a set of rankings. Say you want to study emotional content, you have to go through each report and rank how positive or negative the emotions in a dream were.
Not only can the subjects distort the research by failing to perfectly recall the dream, but human bias during coding is virtually unavoidable.
University of Ottawa’s Joseph De Koninck studies what our minds are busy doing while we sleep. He studies the (usually more negative than positive) emotions of dreaming, and is interested in how these emotions develop throughout the dreams. As a dream psychologist, De Koninck must continually work with human error introduced during the coding process.
What if researchers could eliminate the need for a human judge altogether? Computers can be taught to identify the level of emotion in a written text. This sort of Artificial Intelligence uses algorithms that can be trained from databases of reports and their corresponding rankings by human judges. The computer model uses individual words and the reoccurrence of words throughout the text to shift rankings and take into account words like “not” that flip the meaning.
Most importantly for De Koninck’s research, the computer algorithm can follow the evolution of rankings as dreams progress. By quickly ranking many accounts, it can give statistical information on the evolution of dreamers’ emotions.
The computer program has the possibility to agree with the human judge 65 per cent of the time, and was hardly ever worse than a ranking from the human judge. That’s quite good considering that human judges only agree 60–80 per cent of the time. With such good agreement, these electronic judges could be used to quickly mine the huge number of dream accounts available. De Koninck wouldn’t have to rank each one individually or worry about human bias—and that sounds like a dream come true for scientists.