Suicide Detection in Tweets Using LSTM and Transformers

Twitter is one of the most popular and commonly used social media platforms among people of all ages when textually expressing their state of mind or emotions. Some of these emotions and thoughts are from people who are suicidal. Suicide is a serious public health problem in countries around the world. To detect suicidal thoughts or opinions from users’ tweets, we can examine and classify textual content on Twitter. In this work, we train two deep learning models, Bi-LSTM and transformers including BERT, RoBERTa, DistillBERT, DeBERT, BART and mBART, to classify tweets as either suicidal or otherwise. Our results show that BERT performed best with an accuracy of 98%, while the Bi-LSTM model achieved an accuracy of 95%. We also tested the effectiveness of some pre-trained models achieving 99.6% accuracy with RoBERTa, 99.6% accuracy with DeBERTa, 97.7% accuracy using DistilBERT, 99.9% accuracy with BART, and 99.4% accuracy achieved with mBART. With these results, applying deep learning models to social media can help to timely identify an individual likely to commit suicide based on their tweets and aid the relatives in providing the necessary help.