Detecting and Determining Degree of Suicidal Ideation on Tweets Using LSTM and Machine Learning Models

Abstract

In recent times, Twitter has been examined as a tool for identifying people's mental health state, such as depression and suicidality. Increasingly, those who are depressed or suicidal express themselves on social media. Twitter has established internal reporting systems after realising that individual users indicate suicidality on their platform. (Twitter Inc, 2014). This type of risk detection is not automatic, does not occur in real-time, and relies solely on the discretion of concerned users, of whom many have difficulty determining genuine risk. The objective of this project is to provide automatic early detection of suicidal degree in twitter user profiles by evaluating tweet data. From 26th December 2022 to 10th January 2023, Twitter was monitored for a series of suicide-related phrases and terms using the public API. When tweets that match the criteria were found we further collected the tweets from the timeline of the users. During this time, 16,820 tweets from 529 users were collected: 40% were randomly selected and were manually labelled as ‘1’ for suicidal and ‘0’ for non-suicidal tweets, the rest were used for testing the model. A LSTM model was applied on the data to find suicidal tweets from non-suicidal tweets with an accuracy of 90.8%. A machine learning model was further applied on these tweets and were given one of the following degree ‘Deeply Troubling’ (DT), ‘Possibly Troubling’ (PT) and ‘safe to ignore’(STI).. Assigning a degree can help to cater the necessary services based on the degree. Our project aims to directly identify the real users for intervention.