Word vectors for mental health

It might seem quite far-fetched that a vector of numbers generated by a computer can help cure mental illness, yet this is exactly what my master thesis is about. With the experienced mentorship of Gavagai staff and my university supervisor, I have built a system that can predict whether a social media user manifests symptoms of an eating disorder. And not only that, the system also tries to detect the risk as quickly as possible, because as we all know there is no time to waste when it comes to mental health.

My system is based on deep learning methods and uses a technology where words are represented as vectors of numbers. There are different ways to obtain these vectors and the algorithm you choose influences the amount and type of information that is contained in the vectors. Gavagai’s technology is also based on word vectors, so I figured I could evaluate different types of vectors on the background of early risk prediction on the Internet.

I have now reached the mid-term point of my thesis and the results are quite promising. The system can correctly identify 9 out of 10 at-risk users after having read only a few posts from a person’s history. It does send out a few false alarms, but hey, nobody’s perfect. Further on we will also test a new, powerful technique called contextualized embeddings (the most popular are called ELMo and BERT, if you like Sesame Street references). We hope to get even better results with the help of these muscle-men. Fingers crossed!

Elena Fano
Master thesis student at Gavagai

This website uses cookies to ensure you get the best experience.