Using Word Embeddings to Quantify Ethnic Stereotypes in 12 years of Spanish News

Using Word Embeddings to Quantify Ethnic Stereotypes in 12 years of Spanish News Danielly Sorato Universitat Pompeu Fabra Barcelona, Spain danielly.sorato@upf.edu Diana Zavala-Rojas European Social Survey ERIC Universitat Pompeu Fabra Barcelona, Spain diana.zavala@upf.edu Carme Colominas Ventura Universitat Pompeu Fabra Barcelona, Spain carme.colominas@upf.edu Abstract The current study provides a diachronic anal- ysis of the stereotypical portrayals concerning seven of the most prominent foreign national- ities living in Spain in a Spanish news outlet. We use 12 years (2007-2018) of news articles to train word embedding models to quantify the association of such outgroups with drug use, prostitution, crimes, and poverty concepts. Then, we investigate the effects of sociopoliti- cal variables on the computed bias series, such as the outgroup size in the host country and the rate of the population receiving unemployment benefits. Our findings indicate that the texts exhibit bias against foreign-born people, espe- cially in the case of outgroups for which the country of origin has a lower Gross Domestic Product per capita (PPP) than Spain. 1 Introduction Languages are complex and systematic instruments of communication that reflect the culture of a given population. By studying language, it is possible to observe stereotypes, a type of social bias that is present when discourse about a given group over- looks the diversity of its members and focuses only on a small set of features (Sánchez-Junquera et al., 2021; Tajfel et al., 1964). As such, language analy- sis is a good way to depict, understand, and demon- strate stereotypes (Garg et al., 2018; Basow, 1992; Wetherell and Potter, 1993; Bonilla-Silva and For- man, 2000). Nonetheless, like society, languages are not static. Variations in lexical systems can be observed over time due to a myriad of intra- and extra-linguistic factors. By analyzing extra- linguistic aspects, it is possible to gain insights into the dynamics of social, cultural, and politi- cal phenomena reflected in texts (Marakasova and Neidhardt, 2020). Efficient methods for performing diachronic analysis are crucial, as manually evaluating sev- eral years of text collections is unfeasible due to the large amount of data involved. As such, compu- tational methods for diachronic linguistic analysis are of utmost importance, and ongoing research shows that word embeddings models are helpful tools to this end (Garg et al., 2018; Kroon et al., 2020; Hamilton et al., 2016; Kutuzov et al., 2018; Lauscher et al., 2020). Word embeddings are powerful representations of language, that allow for the quantification of re- lationships between words through efficient numer- ical operations inside the vector space. In this con- text, previous works demonstrated that such models contain machine-learned biases in their geometry that closely depict societal stereotypes (Bolukbasi et al., 2016b; Gonen and Goldberg, 2019; Garg et al., 2018; Kroon et al., 2020), which is not sur- prising since stereotypes are massively present in texts used to train computational models (Sánchez- Junquera et al., 2021; Nadeem et al., 2020). Al- though such language models should be carefully tested for biases and not blindly applied to widely computational applications due to …

Content truncated to 500 words