Suicidal ideation detection using natural language processing

• Main Authors: Gugalia, Charvy, Jindal, Mudit, Rana, Manoj Kumar, Winster, S. Godfrey
• Format: Conference Proceeding
• Language: English
• Subjects: Accuracy; Artificial neural networks; Datasets; Deep learning; Digital media; Machine learning; Natural language processing; Social networks; Suicidal behavior; Suicidal ideation; Suicides & suicide attempts; Target detection
• ISSN: 0094-243X; 1551-7616
• DOI: 10.1063/5.0217127

Suicide is a very critical issue in modern society. To save people’s lives it is essential to identify and prevent the suicide attempts. The most crucial and difficult part is to be able to detect the idea of suicide. Understanding a person’s thoughts and to also understand the factors and signs that might lead to suicide is very important. In many researches, researchers have seen that posts on social media can be used as a criteria to detect any suicidal ideation. Two methods for detecting suicidal ideation are currently being researched: clinical techniques based on interactions between social experts and the target audiance, as well as Machine Learning (ML) technology including feature engineering or Deep Learning (DL) for the purpose of automatic detection based on online contents. We can obtain faster and accurate detection of ideation of suicide using Deep Learning based approaches that are used for classification. By the combination the Convolutional Neural Network (CNN) and Long-Short-Term-Memory (LSTM) models can be employed for this purpose to detect such ideations from the posts of users. Some methods, such as adding more training data, employing attention models to boost the effectiveness of current models, etc., could be used to increase accuracy. In this paper a LSTM and CNN ensembled model is proposed to detect the suicidal ideations by analyzing the social media submissions. The proposed CNN Model has achieved the accuracy of 98.60% on the training dataset and 91.71% on the testing dataset. The proposed LSTM Model has achieved the accuracy of 95.82% on the training dataset and 93.89% on the testing dataset.