Disambiguation of features for improving target class detection from social media text

The rise of social media has led to an abundance of textual data, as well as the rise of unhealthy behaviours targeted at others (e.g. bullying, hate speech) or at oneself (e.g. suicide). In recent years, machine learning approaches have been employed to detect such behaviours, which tend to constitute a small portion of the social media content and need to be distinguished from other discourse on social media that may discuss such behaviours without displaying that behaviour, e.g. social media posts about helping people who may be at risk of suicide, thus, making this a very challenging task. In the context of machine learning, such behaviours are referred to as target classes, i.e. the main behaviours to be detected. In this paper we proposed an approach for disambiguation of features in relation to their membership to the target class vs. non-target class(es). We validate our approach with a case study on suicide detection and our results show that the proposed disambiguation approach leads to a better detection rate of suicide.