Investigating Depression Semantics on Reddit

Major depression is a challenging issue affecting individuals and those of the people around them. This paper investigates the Reddit comments for the automated identification of comments being indicative of depressive behaviour. We measure the socio-psycho-linguistic attributes as useful indicators and their importance for characterising the depression content. We tested content-level classifiers on Reddit data. The proposed BERT and BiLSTM with attention model outperform baseline machine learning (ML) and deep learning (DL) models and achieve a weighted F1-score of 0.81 and 0.84 respectively. Our results reveal that while semi-supervised BERT underperform a few ML models, it still gives non-zero classification and high class-wise precision for non-depressed class.