Time-Aware Predictions of Moments of Change in Longitudinal User Posts on Social Media

Anthony Hills, Adam Tsakalidis, Maria Liakata

Abstract: Capturing changes in an individual’s language is an important aspect of personalised mental health monitoring. A key component is modelling the influence of time, as contextual information both in the recent or distant past/future carries varying semantic weight. We capture and contrast this information by identifying neural, time-sensitive, bi-directional representations of individuals – modelling time-intervals in their social-media posts inspired by the Hawkes process. We demonstrate that our approach helps identify whether an individual’s mood is changing drastically, or smoothly on two social media datasets – yielding superior performance compared to time-insensitive baselines and outperforming the state-of-the-art on the CLPsych 2022 shared task.