Mental Health Classification using Large Scale Tweet Dataset

Abstract

Mental health is socially an important issue in society. In recent years, the mental health issue has been closely related to activities on the Internet. We have developed a system for classifying Twitter users whether they have mental health issues. We assumed that accounts that match specific patterns are likely to have mental health issues, and collected positive and negative examples based on this assumption. In order to investigate the possibility to classify user utterances without mental health specific keywords, we discarded tweets which include such keywords, in contrast to keyword dependent previous studies. We trained and compared classification performances of BERT, LSTM, and SVM. Regarding BERT, we originally pre-trained the model using the large-scale tweet data. BERT achieved high classification performance with Accuracy 0.83, Recall 0.8, Precision 0.88, and F1-score 0.84 in the best performance case.