Advanced deep learning and large language models for suicide ideation detection on social media

Recently, suicide ideations represent a worldwide health concern and pose many anticipation challenges. Actually, the prevalence of expressing self-destructive thoughts especially on forums and social media requires effective monitoring for suicide prevention, and early intervention. Meanwhile, deep learning techniques and Large Language Models (LLMs) have emerged as promising tools in diverse Natural Language Processing (NLP) tasks, including sentiment analysis and text classification. In this paper, we propose a deep learning model incorporating triple models of word embeddings, as well as various fine-tuned LLMs, to identify suicidal thoughts in Reddit posts. In effect, we implemented a Bidirectional Long Short-Term Memory (BiLSTM), and a Convolutional Neural Network (CNN) model to categorize posts associated with non-suicidal and suicidal thoughts. Besides, through the combination of Word2Vec, FastText and GloVe embeddings, our models learn intricate patterns and prevalent nuances in suicide-related language. Furthermore, we employed a merged version of CNN and BiLSTM models, entitled C-BiLSTM, and several LLMs, including pre-trained Bidirectional Encoder Representations from Transformers (BERT) models and a Generative Pre-training Transformer (GPT) model. The analysis of all our proposed models shows that our C-BiLSTM model with triple word embedding and our GPT model got the best performance compared to deep learning and LLMs baseline models, reaching accuracies of 94.5% and 97.69%, respectively. In fact, our best model’s capacity to extract meaningful interdependencies among words significantly promotes its classification performance. This analysis contributes to a deeper understanding of the psychological factors and linguistic markers indicative of suicidal thoughts, thereby informing future research and intervention strategies.