(ICCC 2023) High-Precision Detection of Suicidal Ideation on Social Media Using Bi-LSTM and BERT Models

In a poignant moment on social media, a post suggesting suicidal thoughts is rapidly overshadowed by an array of memes, updates, and ads. Such overlooked cries for help underscore the dire need for reliable detection mechanisms. Addressing this urgent issue, our research introduces an innovative approach to detecting suicidal ideation from text. Using TF-IDF and Chi-square tests, we identify crucial keywords and narrow them down for frequency analysis. Our framework leverages Word2Vec embeddings and is further enriched with readability scores, sentiment assessments via Text-Blob, and text length metrics. Latent Dirichlet Allocation (LDA) aids in topic modeling, its insights visualized through compelling word clouds. This multifaceted approach enhances predictive accuracy and provides a nuanced understanding of suicidal ideation. We initially employed conventional algorithms like AdaBoost, Random Forest, XGBoost, and Logistic Regression before transitioning to advanced deep learning models. The Bi-LSTM and BERT models emerged as top performers, achieving detection accuracies of 97% and 98%, respectively. While our study is constrained by dataset limitations and potential biases, it marks a significant step forward in the identification of suicidal tendencies in digital spaces. Looking for-ward, broader dataset integration, refined models, and a continuous recalibration in line with evolving digital discourse are anticipated next steps.