Automatic Depression Detection
Depression is probably the most widespread mental illness in the world. It affects millions of people and is a huge burden on the health services. Any sort of automatic depression detection would be a wonderful breakthrough.
The way that people speak can suggest whether they are in depression or not. But as long as therapists were just taking notes this knowledge was not very useful. Today, there are millions of people on the Internet, on social media, and in chat rooms. They are all exchanging ideas in text. This means that there is a vast body of text, millions upon millions of words, on every possible subject. Scientists are analysing these texts to try to identify depression. The results are quite surprising.
Machine learning for depression
Language has two parts: content and style. Content is what you talk about, style is how you say it. Analyzing content shows that depressed people use a higher proportion of negative words. These are words such as "sad, lonely, unhappy, miserable, can't cope". Depressed people also have a different focus. Automatic depression detection software shows that depressed people talk much more about "I, me, myself" than they do about "they, them, people". It seems that depressed people are more inward focused, on themselves.
Of course, it is not possible to tell which comes first. Does depression cause you to focus inwards, and use the 'I' word. Or does focusing inwards cause depression?
Depressed people also have a different style of speaking. They tend to speak in absolutes. "Everything, nothing, everybody, never". Research is showing that these 'absolutes' are a better indicator of depression than words about negative emotions.
These results are very encouraging. Automatic depression detection is fast, cheap, and appears to be effective. In fact, in many areas of mental health, automatic analysis is proving faster and more accurate than experts in those fields. As more and more text becomes available for analysing, the results will get even more accurate.