Detecting Signs of Depression from Social Media Texts using Generalized Autoregressive Pretraining Transformer Model

Recognizing signs or symptoms of depression involves determining subtle linguistic cues and patterns within the textual content to reveal indicators of someone’s potential depressive state. Due to the exponential growth of textual content on the web or social media in recent times, it is a challenging task to identify signs of depression. This paper presents a deep learning-based technique to detect signs of depression from social media texts. The critical objective of the work is to develop an intelligent framework that can accurately identify the level of depression from social media posts in English, categorizing them into three classes: not depression, moderate, and severe. This work exploits several machine learning, deep learning, and transformer-based techniques to address the task. Through extensive experimentation and fine-tuning, we evaluated multiple models, including LR, MNB, SVM, CNN, BiLSTM, CNN+BiLSTM, MBERT, XLM-R
base
, BERT
base
, and XLNet
base
. Experimental results demonstrate that the XLNet achieved the highest macro F1-score of 0.43. This result highlights the substantial potential of the XLNet
base
model in effectively identifying and detecting signs of depression from social media text.