Automatic Identification of Suicidal Ideation in Texts Using Cascade Classifiers

According to the 2021 World Health Organization report, suicide is a universal phenomenon that causes about 703,000 deaths per year, being among the first five causes of unnatural death. Suicide mainly affects young people between 15 and 29 years old, who are also the main users of social media. It is noteworthy that these digital platforms play a dual role in suicide issues, on the one hand, by allowing access to dangerous sites that can provide pro-suicide information and, on the other hand, by allowing clues of suicidal ideations to be detected through shared content. To address this health concern, this study presents a computational method based on a cascade classification that first detects the distribution of latent emotions in text and uses this output to identify signs of suicidal ideation. Our experimental results show that the cascade architecture proves to be more robust than direct classification when there are no explicit signs of suicidal ideation. In addition, unlike direct classification, our proposed approach automatically provides information about the emotions that influence a person with suicidal thoughts.