Learning from Mental Disorder Self-tests: Multi-head Siamese Network for Few-shot Knowledge Learning

Anonymous

Abstract: Social media is one of the most highly sought resources to analyze characteristics of the language by its users. In particular, many researchers utilized various linguistic features to identify users with mental disorders. However, generalizing linguistic features of such psychiatric patients is challenging since these features are apparently dependent on cultural or personal language habits. To address this challenge, we make use of the symptoms, which are shared properties of people with mental illness, concerning clinical contents rather than the ways of expressing them. In this paper, we aim to let our classification model identify informative features by training on knowledge about the symptoms. To this end, we propose a multi-head siamese network, which captures informative features based on the knowledge of mental illness symptoms and compares them to those of target text to be classified. The model is designed to learn the required knowledge by reading just a few questions from self-tests, and to identify similar stories from social media texts. Experimental results demonstrate that our model achieves improved performance as well as human-interpretable results for mental illness symptoms. A case study shows that our proposed model offers the possibility of automatic mental illness diagnosis, grounded on rational reasons.

Paper Type: long