The core objective is to develop a reliable and accurate system capable of identifying depressed and non-depressed users through natural language processing (NLP) techniques. To this end, we experimented with a range of sequence modelling approaches, including Simple RNN, Unidirectional and Bidirectional LSTM, Gated Recurrent Units (GRU), and their bidirectional counterparts. Furthermore, we integrated a transfer learning approach using BERT (Bidirectional Encoder Representations from Transformers) to analyse complex sentence structures and contextual relationships in user posts. All models were trained and evaluated on a labelled dataset consisting of social media text entries annotated for depression. The evaluation metric of choice was Recall, given the critical importance of minimizing false negatives in mental health detection. The BERT model achieved the highest recall score of 98.10%, followed closely by Bidirectional LSTM and Bidirectional GRU models, both scoring 97.76%. These results demonstrate the superiority of bidirectional architectures in capturing contextual semantics and underline the importance of transfer-based models like BERT in improving mental health classification performance. Our study confirms that bidirectionality and contextual embedding’s significantly boost detection capabilities in NLP-based mental health applications. This work contributes to building intelligent systems for early diagnosis of depression, supporting mental health professionals with enhanced screening tools in online environments.
Garg et al. (Tue,) studied this question.