Deep learning algorithm for suicide sentiment prediction

Abstract

The increasing use of social media provides unprecedented access to the behaviors, thoughts, feelings and intentions of individuals. We are interested, in this paper, in the detection of notes that express bad feelings that might lead to committing suicide. Our goal is to present an automated detection and prediction system capable of recognizing severe depression through analyzing sentiments and feelings expressed on social networks, blogs, emails and even textual notes. In this work, we have set up a chain of treatments to extract characteristics from notes reflecting the emotional state. We can summarize these treatments in two phases: a pretreatment phase based on the Arabic stemming algorithms, and a phase of construction of feature vectors specific to each word of the corpus based on Term Frequency-Inverse Document Frequency method. Then, we applied a model based on Convolutional Neural Networks to predict the nature of feelings behind the note. The Convolutional Neural Network algorithm is one of many famous algorithms of deep learning field. It is originally created for image processing applications. But recently, it is more and more used in text mining and sentiment analysis field. The originality of the approach is, in one hand, to consider both the nature of the words that individuals used to express themselves. And in the other hand, to use the advantages of the Convolutional Neural Network to automatically extract the most significant and reliable features. A preliminary experiment allowed us to evaluate our approach on real cases of online suicidal notes.