Collection Methods and Data Characteristics of the PagkataoKo Dataset

Collection Methods and Data Characteristics of the PagkataoKo Dataset Edward Tighe, Luigi Acorda, Alexander Agno II, Jesah Gano, Timothy Go, Gabriel Santiago, Claude Sedillo De La Salle University, Manila, Philippines {edward.tighe, luigi acorda, alexander agno, jesah gano, timothy go, gabriel santiago, claude sedillo}@dlsu.edu.ph Abstract We present the PagkataoKo Dataset, a new dataset for Filipino Automatic Personal- ity Recognition (APR) containing demo- graphic, personality trait, and social media data sourced from 3,128 Filipino Instagram and/or Twitter users. As APR is focused on processing an individual’s observable ac- tions, we improve upon the previous Fil- ipino APR dataset by collecting multimodal data, as well as similar data expressed in dif- ferent environments. In our paper, we de- scribe our collection methodology and de- tail the general characteristics of the dataset. We also report the language characteristics of the posts and highlight the presence of multiple languages (e.g. English, Tagalog) and code-switching – two aspects that make Filipino APR difficult. Lastly, we discuss how our dataset provides future work with multiple options to explore in order to nav- igate around the complexities found in Fil- ipinos’ language usage. 1 Introduction Automatic Personality Recognition (APR) is a computing task that focuses on personality traits (i.e. individual differences) and their externaliza- tions (Vinciarelli and Mohammadi, 2014). The task is rooted in the idea that traits influence a person’s interactions in different environments (Larsen and Buss, 2008); hence, a person’s ob- servable actions contain traces of their traits, such as how it is believed that important personality characteristics and individual differences are en- coded into one’s language (i.e. the lexical hypoth- esis) (Goldberg, 1981; Tausczik and Pennebaker, 2010). As a computing task, APR, therefore, in- volves the collection and processing of these ob- servable actions and analysis (e.g. descriptive, causal, predictive) of any traces of personality left behind. APR has received much attention over the past two decades – leading to the exploration of many different types of mediums (e.g. text, image, audio) in which personality may have been ex- pressed upon. Early work in APR mostly fo- cused on studying language usage – with data coming from conversation recordings (Mehl et al., 2006; Mairesse et al., 2007), emails (Gill and Oberlander, 2002), and essays (Pennebaker and King, 1999; Argamon et al., 2005; Mairesse et al., 2007). These early studies did not produce high-performing predictive models nor did they have a high volume of data to validate results with but were at least able to show that indica- tors of personality can be found in one’s language. APR studies then branched out and explored other sources of observable actions with a vast number of studies gravitating towards social media plat- forms, such as (but not limited to) Facebook (Gol- beck et al., 2011b; Gosling et al., 2011; Wald et al., 2012; Markovikj et al., 2013; Schwartz et al., 2013; Kosinski et al., 2014; Park et al., 2015; Se- galin et al., 2017), Twitter (Golbeck et al., 2011a; Quercia et al., 2011; Rangel Pardo et al., 2015; …

Content truncated to 500 words