Hybrid transformer-based approaches for suicidal and depressive posts detection

What is it about?

Depression is a global issue that affects countries and social groups worldwide. To mitigate its impact, it is crucial to provide solutions capable of identifying at-risk person. Nowadays, social networks have become an important source of information that can help detect individuals suffering from depression and who may commit suicide attempts. Therefore, to address this pressing issue, we propose multiple hybrid approaches based on several combinations of six pre-trained language models (BERT, RoBERTa, MentalBERT, MentalRoBERTa, DistilBERT and DistilRoBERTa.) with convolutional neural networks and bidirectional long short term memory to identify suicide risk at post level. We evaluate our approaches on a dataset of 58,000 Reddit posts by conducting a series of comparative experiments using individual PLM models, four traditional machine learning models, and five existing studies that use the same dataset. A comprehensive evaluation of these architectures is performed across a range of metrics and criteria. The results demonstrated that the integration of a CNN or a BiLSTM layer improved the models' performance. The highest F1-score (97.66%) was achieved by both MentalBERT-CNN and MentalBERT-BiLSTM across multiple random seed evaluations.

Featured Image