Abstract

Suicide is a leading cause of death worldwide, particularly among the young generation. The increase in suicidal posts on social media platforms such as Reddit has presented both challenges and opportunities for mental health intervention. Our work aims to use vast data generated by individuals on Reddit by using advanced deep learning techniques to identify suicidal posts and non-suicidal posts. The dataset used is collected from two different datasets: one dataset curated from the Reddit API called (PRAW) with the help of various subreddits (e.g., SuicideWatch, Anxiety, and Depression) and neutral topics (e.g., Jokes, Movies, Popular, Books) and another publicly available dataset from Kaggle. The proposed model uses long-short-term memory LSTM, bidirectional LSTM (Bi-LSTM), gated recurrent units (GRU), bidirectional GRU (Bi-GRU), and modified BERT-based transformers. The BERT-based model performed better in both datasets compared to other models with an accuracy of 98%, a precision of 98. 5%, and a recall of 98. 5%. These experimental results successfully verify the theoretical efficiency and adaptability of the proposed model in real-time suicidal post-detection.

Authors

Swati Bansode, Vaishali Hirlekar, Shashikant Radke
Shah and Anchor Kutchhi Engineering, India

Keywords

Suicidal Post-Detection, Deep Learning, Natural Language Processing, Reddit, BERT, LSTM, Bi-LSTM, GRU and Bi-GRU