Identification of Suicidal Tendencies of Individuals based on the Quantitative Analysis of their Internet Texts

Abstract

Even though suicide is one of the top three causes of young people’s deaths, no reliable methods of identifying suicidal behavior have been developed. One of the promising directions of research is quantitative analysis of speech. It is nowadays common to process texts by suicidal individuals (mostly suicidal notes or literary texts by famous people, e.g., poets, writes, etc.) and texts by individuals from a control group using software (mostly LIWC) and to design models for classifying texts as those by suicidal individuals or not. This kind of analysis has been mainly performed for English texts that generally have a number of restrictions due to their linguistic nature. The authors are the first to attempt to design a mathematical model to classify texts as those by suicidal or nonsuicidal individuals using numerical values of linguistic parameters as features. Texts (blogs by young people who committed suicides, similar in both genre and topic, to those by individuals of an age-corresponding control group) were processed using the Russian version of LIWC with users’ dictionaries. Unlike current studies, in designing the model we mostly made use of features that are not significantly dependent on the content. This is because not all individuals who committed suicides are known to deal with the topic in their texts. The resulting model was shown to be 71.5% accurate, which is comparable with the state-of-the-art for English texts.