"TweetGenie: automatic age prediction from tweets" by D. Nguyen, R. Gravel, D. Trieschnigg, and T. Meder; with Ching-man Au Yeung as coordinator

Abstract

A person's language use reveals much about the person's social identity, which is based on the social categories a person belongs to including age and gender. We discuss the development of TweetGenie, a computer program that predicts the age of Twitter users based on their language use. We explore age prediction in three different ways: classifying users into age categories, by life stages, and predicting their exact age. An automatic system achieves better performance than humans on these tasks. Both humans and the automatic systems tend to underpredict the age of older people. We find that most linguistic changes occur when people are young, and that after around 30 years the studied variables show little change, making it difficult to predict the ages of older Twitter users.

The coming of age: How do linguists tease apart chronological, biological and social age?

“…For example, in Nguyen et al. (2013), chronological age was a better predictor of the linguistic phenomena in question for younger speakers, while for older speakers, chronological age did not yield clear results. Such findings point to the question of whether chronological age is in fact a good correlate of socially conditioned language change, as is often implicitly assumed in language change studies (both apparent‐ and real‐time ones; see Sankoff, 2006).…”

“…As pointed out in the studies cited above, as well as by Eckert (1996), the vast majority of experimental work generally views a speaker's date of birth, or chronological age , as a reflection of their social age , that is, the age‐related social meanings and roles claimed by the speaker (for a discussion of the term ‘social age’, see Clark‐Kazak (2009). Nevertheless, as shown by studies such as Dubois and Horvath (1999) and Nguyen, Gravel, Trieschnigg, and Meder (2013), chronological age does not necessarily correlate strongly with social age. Speakers may feel and act a different age than what their birth certificates would predict—and they might also not.…”

“…One way to frame such evidence is through asking participants to name or discuss various aspects of their lifespan, defined as a ‘set of culturally specific and linguistically ascribed categories’ (Coupland & Ylanne‐McEwen, 1993). Operationalisations of the lifespan from various fields include age categories (e.g., teenager, young adult, adult, middle aged, old or very old, see e.g., Markides & Boldt, 1983), age groups (e.g., members of the third or fourth ages—Laslett, 1991; and the ‘young‐old’ and the ‘old‐old’—Neugarten, 1974) or life stages (e.g., school pupil, high schooler, student, worker, pensioner, see e.g., Nguyen et al., 2013). 4 In including such questions in interviews, however, the researcher needs to be mindful of their limitations.…”

Mining Social Media Data to Study the Consequences of Dementia Diagnosis on Caregivers and Relatives

“…The algorithm we used to automatically determine the age and sex of authors was developed using Facebook data [14], was further evaluated on blogs and tweets and we judged this tool to be suitable to predict age and sex of the authors in our Reddit post dataset. However, it has been recognised that using textual content from online media to predict age can be insufficient [20].…”

Real Men Don’t Say “Cute”

“…In computer science, there has been growing interest in automatically identifying author traits from their text using machine learning techniques, predicting gender (Burger, Henderson, King, & Zarrella, 2011), age (Rao et al, 2011; Sap et al, 2014), political orientation (Pennacchiotti & Popescu, 2011), and income (Preoţiuc-Pietro, Volkova, Lampos, Bachrach, & Aletras, 2015). Other studies have predicted language associations with perceived author traits for demographic characteristics such as gender (Nguyen et al, 2014), age (Nguyen, Gravel, Trieschnigg, & Meder, 2013), and other features. In these cases, researchers employed raters to categorize author profiles according to their perceptions of the author’s traits.…”