Gender, Age, and Depression detection with BERT-based multitask model

The Author Profiling (AP) task mainly focuses on predicting the demographic and the socio-cultural information of a particular author from his writing; this information could be the gender, age, language variety, religion, etc. Meanwhile, the Depression Detection (DP) task generally aims to identify the depressive content, by classifying a textual document in two categories: depressive or not. The works in the literature treat the two tasks separately, as they construct a model for predicting each information solely; Instead in this work, we propose a new multi-task model based on BERT to predict simultaneously the gender and the age of the author from his publication and the depressive publications. For the training of the model, we have constructed a dataset with publications from the Reddit platform published within 2005-06 and 2017-02 timespan, by adopting an automatic approach for the author gender and age annotation based on Regular Expression (REGEX), and the depression label we considered the publications in a preselected depressive subreddits. Regarding our approach performances, it has demonstrated a competitive result with the State Of The Art (SOTA) models for AP and DP tasks. The results of our model were 0.72, 0.50, and 0.92 with the accuracy metric on the gender, age, and depression detection respectively.