Online suicide ideation detection (OnSIDe): a context-aware transfer learning approach using BERT-CNN

Background Online suicide ideation is a global concern because of its anonymity and reach, and while automated detection using technologies like natural language processing and machine learning holds promise, accurately distinguishing between genuine distress and non-harmful discussions remains a challenge. This study developed an online suicide ideation detection model using a context-aware transfer learning deep learning approach. Method The bidirectional encoder representations from transformers (BERT), a pre-trained context-aware language model, was used with a convolutional neural network (CNN) to develop our proposed model, OnSIDe bidirectional encoder representations from Transformers-convolutional neural network (BERT-CNN). The model was trained, validated, and tested using an English dataset sourced from Reddit ( n = 175,975), and subsequently evaluated on a previously unseen dataset from Twitter ( n = 9,119). Results Experiments revealed OnSIDe BERT-CNN to outperform the baseline models, with an accuracy of 93.25% and an F-score of 92.88%. A reliable performance was also noted for the Twitter dataset (accuracy = 83.77%; F-score = 83.81%). Limitation OnSIDe BERT-CNN was solely trained and tested on English datasets. Conclusion A context-aware suicide ideation detection model is more accurate and impactful by considering the specific textual nuances surrounding a person’s expressed distress.