Author Profiling from Facebook Corpora

Author Profiling from Facebook Corpora Fernando Chiu Hsieh, Rafael Felipe Sandroni Dias, Ivandr´e Paraboni University of S˜ao Paulo, School of Arts, Sciences and Humanities S˜ao Paulo, Brazil fernando.hsieh@usp.br , rafaelsandronidias@gmail.com, ivandre@usp.br Abstract Author profiling - the computational task of prediction author’s demographics from text - has been a popular research topic in the NLP field, and also the focus of a number of prominent shared tasks. Author profiling is a problem of growing importance, with applications in forensics, security, sales and many others. In recent years, text available from social networks has become a primary source for computational models of author profiling, but existing studies are still largely focused on age and gender prediction, and are in many cases limited to the use of English text. Other languages, and other author profiling tasks, remain somewhat less popular. As a means to further this issue, in this work we present initial results of a number of author profiling tasks from a Facebook corpus in the Brazilian Portuguese language. As in previous studies, our own work will focus on both standard gender and age prediction tasks but, in addition to these, we will also address two less usual author profiling tasks, namely, predicting an author’s degree of religiosity and IT background status. The tasks are modelled by making use of different knowledge sources, and results of alternative approaches are discussed. Keywords: Author profiling, Document Classification, Corpus-based approaches 1. Introduction Author profiling - the computational task of prediction au- thor’s demographics from text - has been a popular research topic in the NLP field, and also the focus of a number of prominent shared tasks (Pardo et al., 2017). Author profil- ing is a problem of growing importance, with applications in forensics, security, sales and many others (Rangel et al., 2015). In recent years, text available from social networks has be- come a primary source of data for computational author profiling. However, existing studies of this kind are of- ten focused on age and gender prediction, and are in many cases limited to the use of English text and a few others. Other languages, and other author profiling tasks, by con- trast, remain somewhat less popular. In this work we address the issue of computational au- thor profiling from a Facebook corpus in the Brazilian Por- tuguese language. In doing so, we address standard age and gender prediction tasks, and also two less-known al- ternatives: predicting an author’s degree of religiosity and his/her IT background status. Religiosity prediction may be seen as an extension of personality recognition (Mairesse et al., 2007), and degrees of religiosity are indeed known to correlate with certain personality traits in the Big Five model (de Andrade, 2008). IT background status predic- tion, on the other hand, may be seen as a means to group authors into (e.g., professional) communities. In order to predict age, gender, religiosity and IT back- ground status, we propose a number of author profile models based on standard document classification meth- ods (from bag …

Content truncated to 500 words