Predicting Mental-Illness from Twitter Activity Using Activity Theory Based Context Ontology

Abstract

In this digital era, people are becoming more comfortable to express their sentiments on social media rather than sharing them in person. The analysis of these sentiments is a doorway to understand their mental-health status. The literature proposed linguistic features-based methods for their automatic analysis to see if the user needs any psychological assistance. However, these methods analysed expressed sentiments as discrete and fragmented information, without considering the holistic context. For this, activity theory has been employed to explicate the context of sentiment expression activity of a Twitter user. This contextual information of tweeting activity is utilised for better classifier design in order to predict mental-illness. During the first step, tweets of self-reported mentally-ill users are collected. Then, the contextual information from Twitter user account and tweets is populated into activity theory-based context ontology. The activity components and their associated fields are used as a contextual feature-set for SVM training. The solution is tested by a prototype application, on tweets of self-reported mentally-ill users, and the results of the proposed contextual features are compared with classic linguistic features. The proposed features increased the accuracy by 23.11%, which proves the hypothesis that the inclusion of context is valuable for sentiment analysis.