Combining Heterogeneous User Generated Data to Sense Well-being

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 3007–3018, Osaka, Japan, December 11-17 2016. Combining Heterogeneous User Generated Data to Sense Well-being Adam Tsakalidis University of Warwick Coventry, UK a.tsakalidis@warwick.ac.uk Maria Liakata University of Warwick Coventry, UK m.liakata@warwick.ac.uk Theo Damoulas University of Warwick Coventry, UK t.damoulas@warwick.ac.uk Brigitte Jellinek Salzburg University of Applied Sciences FH Salzburg, Puch Urstein ¨Osterreich brigitte.jellinek@fh-salzburg.ac.at Weisi Guo University of Warwick Coventry, UK weisi.guo@warwick.ac.uk Alexandra I. Cristea University of Warwick Coventry, UK a.i.cristea@warwick.ac.uk Abstract In this paper we address a new problem of predicting affect and well-being scales in a real-world setting of heterogeneous, longitudinal and non-synchronous textual as well as non-linguistic data that can be harvested from on-line media and mobile phones. We describe the method for collect- ing the heterogeneous longitudinal data, how features are extracted to address missing informa- tion and differences in temporal alignment, and how the latter are combined to yield promising predictions of affect and well-being on the basis of widely used psychological scales. We achieve a coefficient of determination (R2) of 0.71 − 0.76 and a ρ of 0.68 − 0.87 which is higher than the state-of-the art in equivalent multi-modal tasks for affect. 1 Introduction The World Health Organisation describes mental health as “the foundation for well-being and effective functioning for an individual and for a community” and highlights the importance of selecting suitable indicators of mental health (Herrman et al., 2005). One can distinguish between macro-level indicators, which are meant to provide a picture of generic well-being across a large population, usually at national scale, and individual indicators of mental health. Most of the macro measures typically use statistics from census, administrative and economic sources to measure the social and economic macro-environment as important determinants of mental health (e.g. Human Development Index, Gender Development Index, Human Poverty Indices (OECD, 2013). With the advent of widely available social media data, there have also been efforts to automatically obtain macro indicators of well-being and happiness, primarily through the analysis of geolocated Twitter posts (Dodds et al., 2011; Lansdall-Welfare et al., 2012; Lampos et al., 2013). These pieces of work seek to identify occurrence patterns for words with pre-defined affect scores at different levels of temporal granularity. Such approaches, with more sophisticated components for emotion recognition in social media content, can be alternatives to public surveys for mood and happiness indicators. At the other end of the spectrum we have individual indicators of mental health. These include mea- sures of positive mental health, such as coherence & meaning in life, self-esteem etc. as well as indicators of mental distress, such as negativity, anxiety, depression (Herrman et al., 2005). These measures can be used by experts or individuals for diagnostic and management purposes, but also in aggregation, for large scale surveys. However, the reliance on self-reporting required to obtain these measures is time consum- ing and expensive and can only produce sparse data on small populations. Moreover, self-reporting is likely to introduce bias into results. Recent …

Content truncated to 500 words