A comparative analysis of transfer learning models on suicide and non-suicide textual data

Abstract

The rise of social media has allowed individuals to express themselves freely, increasing the visibility of mental health concerns, including suicidal tendencies. This issue is particularly significant, as suicide is one of the leading causes of death globally. The objective of this study is to develop a model capable of accurately detecting suicide-related textual data using advanced natural language processing techniques. To achieve this, we applied transfer learning models, including bidirectional encoder representations from transformers (BERT), robustly optimized bidirectional encoder representations from transformers (RoBERT), a lite BERT (ALBERT), and decoding-enhanced BERT with disentangled attention (DeBERTa). the dataset used in this research includes 232,074 posts from Reddit, categorized into suicide and non-suicide labels. Preprocessing steps such as removing HTML tags, special characters, and punctuation were applied, followed by stopword removal and lemmatization. The models were trained and evaluated using accuracy, precision, recall, and F1-score metrics. Among the models tested, DeBERTa demonstrated superior performance, achieving an accuracy of 98.70% and an F1-score of 98.70%. These findings suggest that transfer learning models, particularly DeBERTa, are effective in identifying suicidal ideation in textual data.